diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 5d6a2bb..2e5f357 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -11,12 +11,11 @@ on: types: [opened, synchronize, reopened] jobs: - test: - runs-on: ubuntu-latest - timeout-minutes: 60 + run-tests: + runs-on: ubuntu-22.04 strategy: matrix: - python-version: [3.8, 3.9, '3.10', '3.11'] + python-version: [3.8, 3.9, '3.10', 3.11] steps: - uses: actions/checkout@v2 @@ -48,10 +47,36 @@ jobs: run: | echo "Running tests with Python ${{ matrix.python-version }}" python -m unittest discover -s tests + + check-code-format: + runs-on: ubuntu-22.04 + strategy: + matrix: + python-version: [3.8, 3.9, '3.10', 3.11] - build-and-push: - needs: test - runs-on: ubuntu-latest + steps: + - uses: actions/checkout@v2 + + - name: Set up Python ${{ matrix.python-version }} + uses: actions/setup-python@v2 + with: + python-version: ${{ matrix.python-version }} + + - name: Install dependencies + run: | + python -m pip install --upgrade pip + python -m pip install flake8 + + - name: Lint with flake8 + run: | + # stop the build if there are Python syntax errors or undefined names + flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics + # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide + flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --statistics + + publish-to-pypi: + needs: [run-tests, check-code-format] + runs-on: ubuntu-22.04 if: github.event_name == 'push' && startsWith(github.ref, 'refs/tags') steps: - uses: actions/checkout@v2 diff --git a/run_server.py b/run_server.py index 3feeec6..4c6403b 100644 --- a/run_server.py +++ b/run_server.py @@ -4,15 +4,15 @@ from whisper_live.server import TranscriptionServer if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument('--port', '-p', - type=int, + type=int, default=9090, help="Websocket port to run the server on.") parser.add_argument('--backend', '-b', - type=str, - default='faster_whisper', + type=str, + default='faster_whisper', help='Backends from ["tensorrt", "faster_whisper"]') parser.add_argument('--faster_whisper_custom_model_path', '-fw', - type=str, default=None, + type=str, default=None, help="Custom Faster Whisper Model") parser.add_argument('--trt_model_path', '-trt', type=str, @@ -30,7 +30,7 @@ if __name__ == "__main__": server = TranscriptionServer() server.run( "0.0.0.0", - port=args.port, + port=args.port, backend=args.backend, faster_whisper_custom_model_path=args.faster_whisper_custom_model_path, whisper_tensorrt_path=args.trt_model_path, diff --git a/setup.py b/setup.py index bc00ff1..652affb 100644 --- a/setup.py +++ b/setup.py @@ -10,36 +10,38 @@ HERE = pathlib.Path(__file__).parent README = (HERE / "README.md").read_text() # This call to setup() does all the work -setup(name="whisper-live", - version=__version__, - description="A nearly-live implementation of OpenAI's Whisper.", - long_description=README, - long_description_content_type="text/markdown", - include_package_data=True, - url="https://github.com/collabora/WhisperLive", - author="Collabora Ltd", - author_email="vineet.suryan@collabora.com", - license="MIT", - classifiers=[ - "Development Status :: 4 - Beta", - "Intended Audience :: Developers", - "Intended Audience :: Science/Research", - "License :: OSI Approved :: MIT License", - "Programming Language :: Python :: 3", - "Programming Language :: Python :: 3 :: Only", - "Programming Language :: Python :: 3.8", - "Programming Language :: Python :: 3.9", - "Topic :: Scientific/Engineering :: Artificial Intelligence", - ], - packages=find_packages( - exclude=("examples", - "Audio-Transcription-Chrome", - "Audio-Transcription-Firefox", - "requirements", - "whisper-finetuning" - ) - ), - install_requires=[ +setup( + name="whisper-live", + version=__version__, + description="A nearly-live implementation of OpenAI's Whisper.", + long_description=README, + long_description_content_type="text/markdown", + include_package_data=True, + url="https://github.com/collabora/WhisperLive", + author="Collabora Ltd", + author_email="vineet.suryan@collabora.com", + license="MIT", + classifiers=[ + "Development Status :: 4 - Beta", + "Intended Audience :: Developers", + "Intended Audience :: Science/Research", + "License :: OSI Approved :: MIT License", + "Programming Language :: Python :: 3", + "Programming Language :: Python :: 3 :: Only", + "Programming Language :: Python :: 3.8", + "Programming Language :: Python :: 3.9", + "Topic :: Scientific/Engineering :: Artificial Intelligence", + ], + packages=find_packages( + exclude=( + "examples", + "Audio-Transcription-Chrome", + "Audio-Transcription-Firefox", + "requirements", + "whisper-finetuning" + ) + ), + install_requires=[ "PyAudio", "faster-whisper==0.10.0", "torch", @@ -53,6 +55,6 @@ setup(name="whisper-live", "openai-whisper", "kaldialign", "soundfile", - ], - python_requires=">=3.8" + ], + python_requires=">=3.8" ) diff --git a/tests/jfk.flac b/tests/jfk.flac deleted file mode 100644 index e44b7c1..0000000 Binary files a/tests/jfk.flac and /dev/null differ diff --git a/tests/test_client.py b/tests/test_client.py index 2afaba1..468b5a1 100644 --- a/tests/test_client.py +++ b/tests/test_client.py @@ -4,7 +4,8 @@ import scipy import websocket import unittest from unittest.mock import patch, MagicMock -from whisper_live.client import TranscriptionClient, resample +from whisper_live.client import TranscriptionClient +from whisper_live.utils import resample class BaseTestCase(unittest.TestCase): @@ -68,7 +69,7 @@ class TestClientCallbacks(BaseTestCase): ] }) self.client.on_message(self.mock_ws_app, message) - + # Assert that the transcript was updated correctly self.assertEqual(len(self.client.transcript), 2) self.assertEqual(self.client.transcript[1]['text'], "Test transcript 2") @@ -79,14 +80,14 @@ class TestClientCallbacks(BaseTestCase): self.client.on_close(self.mock_ws_app, close_status_code, close_msg) self.assertFalse(self.client.recording) - self.assertFalse(self.client.server_error) + self.assertFalse(self.client.server_error) self.assertFalse(self.client.waiting) - + def test_on_error(self): error_message = "Test Error" self.client.on_error(self.mock_ws_app, error_message) - self.assertTrue(self.client.server_error) + self.assertTrue(self.client.server_error) self.assertEqual(self.client.error_message, error_message) @@ -95,10 +96,10 @@ class TestAudioResampling(unittest.TestCase): original_audio = "assets/jfk.flac" expected_sr = 16000 resampled_audio = resample(original_audio, expected_sr) - + sr, _ = scipy.io.wavfile.read(resampled_audio) self.assertEqual(sr, expected_sr) - + os.remove(resampled_audio) diff --git a/tests/test_server.py b/tests/test_server.py index 34c1288..cd14bb3 100644 --- a/tests/test_server.py +++ b/tests/test_server.py @@ -6,6 +6,8 @@ from unittest import mock import numpy as np import evaluate + +from websockets.exceptions import ConnectionClosed from whisper_live.server import TranscriptionServer from whisper_live.client import TranscriptionClient from whisper.normalizers import EnglishTextNormalizer @@ -14,32 +16,31 @@ from whisper.normalizers import EnglishTextNormalizer class TestTranscriptionServerInitialization(unittest.TestCase): def test_initialization(self): server = TranscriptionServer() - self.assertEqual(server.max_clients, 4) - self.assertEqual(server.max_connection_time, 600) - self.assertDictEqual(server.clients, {}) - self.assertDictEqual(server.websockets, {}) - self.assertDictEqual(server.clients_start_time, {}) + self.assertEqual(server.client_manager.max_clients, 4) + self.assertEqual(server.client_manager.max_connection_time, 600) + self.assertDictEqual(server.client_manager.clients, {}) + self.assertDictEqual(server.client_manager.start_times, {}) class TestGetWaitTime(unittest.TestCase): def setUp(self): self.server = TranscriptionServer() - self.server.clients_start_time = { + self.server.client_manager.start_times = { 'client1': time.time() - 120, 'client2': time.time() - 300 } - self.server.max_connection_time = 600 + self.server.client_manager.max_connection_time = 600 def test_get_wait_time(self): - expected_wait_time = (600 - (time.time() - self.server.clients_start_time['client2'])) / 60 - print(self.server.get_wait_time(), expected_wait_time) - self.assertAlmostEqual(self.server.get_wait_time(), expected_wait_time, places=2) + expected_wait_time = (600 - (time.time() - self.server.client_manager.start_times['client2'])) / 60 + print(self.server.client_manager.get_wait_time(), expected_wait_time) + self.assertAlmostEqual(self.server.client_manager.get_wait_time(), expected_wait_time, places=2) + - class TestServerConnection(unittest.TestCase): def setUp(self): self.server = TranscriptionServer() - + @mock.patch('websockets.WebSocketCommonProtocol') def test_connection(self, mock_websocket): mock_websocket.recv.return_value = json.dumps({ @@ -50,7 +51,6 @@ class TestServerConnection(unittest.TestCase): }) self.server.recv_audio(mock_websocket, "faster_whisper") - @mock.patch('websockets.WebSocketCommonProtocol') def test_recv_audio_exception_handling(self, mock_websocket): mock_websocket.recv.side_effect = [json.dumps({ @@ -58,12 +58,12 @@ class TestServerConnection(unittest.TestCase): 'language': 'en', 'task': 'transcribe', 'model': 'tiny.en' - }), np.array([1, 2, 3]).tobytes()] - + }), np.array([1, 2, 3]).tobytes()] + with self.assertLogs(level="ERROR"): self.server.recv_audio(mock_websocket, "faster_whisper") - - self.assertNotIn(mock_websocket, self.server.clients) + + self.assertNotIn(mock_websocket, self.server.client_manager.clients) class TestServerInferenceAccuracy(unittest.TestCase): @@ -71,12 +71,12 @@ class TestServerInferenceAccuracy(unittest.TestCase): def setUpClass(cls): cls.server_process = subprocess.Popen(["python", "run_server.py"]) # Adjust the command as needed time.sleep(2) - + @classmethod def tearDownClass(cls): cls.server_process.terminate() cls.server_process.wait() - + @mock.patch('pyaudio.PyAudio') def setUp(self, mock_pyaudio): self.mock_pyaudio = mock_pyaudio.return_value @@ -84,16 +84,16 @@ class TestServerInferenceAccuracy(unittest.TestCase): self.mock_pyaudio.open.return_value = self.mock_stream self.metric = evaluate.load("wer") self.normalizer = EnglishTextNormalizer() - self.client = TranscriptionClient( + self.client = TranscriptionClient( "localhost", "9090", model="base.en", lang="en", ) - + def test_inference(self): gt = "And so my fellow Americans, ask not, what your country can do for you. Ask what you can do for your country!" self.client("assets/jfk.flac") with open("output.srt", "r") as f: lines = f.readlines() - prediction = " ".join([l.strip() for l in lines[2::4]]) + prediction = " ".join([line.strip() for line in lines[2::4]]) prediction_normalized = self.normalizer(prediction) gt_normalized = self.normalizer(gt) @@ -103,3 +103,35 @@ class TestServerInferenceAccuracy(unittest.TestCase): references=[gt_normalized] ) self.assertLess(wer, 0.05) + + +class TestExceptionHandling(unittest.TestCase): + def setUp(self): + self.server = TranscriptionServer() + + @mock.patch('websockets.WebSocketCommonProtocol') + def test_connection_closed_exception(self, mock_websocket): + mock_websocket.recv.side_effect = ConnectionClosed(1001, "testing connection closed") + + with self.assertLogs(level="INFO") as log: + self.server.recv_audio(mock_websocket, "faster_whisper") + self.assertTrue(any("Connection closed by client" in message for message in log.output)) + + @mock.patch('websockets.WebSocketCommonProtocol') + def test_json_decode_exception(self, mock_websocket): + mock_websocket.recv.return_value = "invalid json" + + with self.assertLogs(level="ERROR") as log: + self.server.recv_audio(mock_websocket, "faster_whisper") + self.assertTrue(any("Failed to decode JSON from client" in message for message in log.output)) + + @mock.patch('websockets.WebSocketCommonProtocol') + def test_unexpected_exception_handling(self, mock_websocket): + mock_websocket.recv.side_effect = RuntimeError("Unexpected error") + + with self.assertLogs(level="ERROR") as log: + self.server.recv_audio(mock_websocket, "faster_whisper") + for message in log.output: + print(message) + print() + self.assertTrue(any("Unexpected error: Unexpected error" in message for message in log.output)) diff --git a/tests/test_vad.py b/tests/test_vad.py index 8ca7ee3..cfc2d3a 100644 --- a/tests/test_vad.py +++ b/tests/test_vad.py @@ -1,14 +1,12 @@ import unittest import numpy as np -import torch -import scipy.io as sio from whisper_live.tensorrt_utils import load_audio -from whisper_live.vad import VoiceActivityDetection +from whisper_live.vad import VoiceActivityDetector class TestVoiceActivityDetection(unittest.TestCase): def setUp(self): - self.vad = VoiceActivityDetection() + self.vad = VoiceActivityDetector() self.sample_rate = 16000 def generate_silence(self, duration_seconds): @@ -19,10 +17,10 @@ class TestVoiceActivityDetection(unittest.TestCase): def test_vad_silence_detection(self): silence = self.generate_silence(3) - speech_prob = self.vad(torch.from_numpy(silence.copy()), self.sample_rate).item() - self.assertLess(speech_prob, 0.5, "VAD incorrectly identified silence as speech.") + is_speech_present = self.vad(silence.copy()) + self.assertFalse(is_speech_present, "VAD incorrectly identified silence as speech.") def test_vad_speech_detection(self): - audio_tensor = torch.from_numpy(load_audio("assets/jfk.flac")) - speech_prob = self.vad(audio_tensor, self.sample_rate).item() - self.assertGreater(speech_prob, 0.5, "VAD failed to identify speech segment.") \ No newline at end of file + audio_tensor = load_audio("assets/jfk.flac") + is_speech_present = self.vad(audio_tensor) + self.assertTrue(is_speech_present, "VAD failed to identify speech segment.") diff --git a/whisper_live/__version__.py b/whisper_live/__version__.py index 40692a7..3dc1f76 100644 --- a/whisper_live/__version__.py +++ b/whisper_live/__version__.py @@ -1 +1 @@ -__version__="0.1.0" +__version__ = "0.1.0" diff --git a/whisper_live/client.py b/whisper_live/client.py index 15850a0..7a8201c 100644 --- a/whisper_live/client.py +++ b/whisper_live/client.py @@ -2,68 +2,14 @@ import os import wave import numpy as np -import scipy -import ffmpeg import pyaudio import threading -import textwrap import json import websocket import uuid import time - - -def format_time(s): - """Convert seconds (float) to SRT time format.""" - hours = int(s // 3600) - minutes = int((s % 3600) // 60) - seconds = int(s % 60) - milliseconds = int((s - int(s)) * 1000) - return f"{hours:02}:{minutes:02}:{seconds:02},{milliseconds:03}" - -def create_srt_file(segments, output_file): - with open(output_file, 'w', encoding='utf-8') as srt_file: - segment_number = 1 - for segment in segments: - start_time = format_time(float(segment['start'])) - end_time = format_time(float(segment['end'])) - text = segment['text'] - - srt_file.write(f"{segment_number}\n") - srt_file.write(f"{start_time} --> {end_time}\n") - srt_file.write(f"{text}\n\n") - - segment_number += 1 - - -def resample(file: str, sr: int = 16000): - """ - # https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/audio.py#L22 - Open an audio file and read as mono waveform, resampling as necessary, - save the resampled audio - - Args: - file (str): The audio file to open - sr (int): The sample rate to resample the audio if necessary - - Returns: - resampled_file (str): The resampled audio file - """ - try: - # This launches a subprocess to decode audio while down-mixing and resampling as necessary. - # Requires the ffmpeg CLI and `ffmpeg-python` package to be installed. - out, _ = ( - ffmpeg.input(file, threads=0) - .output("-", format="s16le", acodec="pcm_s16le", ac=1, ar=sr) - .run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True) - ) - except ffmpeg.Error as e: - raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e - np_buffer = np.frombuffer(out, dtype=np.int16) - - resampled_file = f"{file.split('.')[0]}_resampled.wav" - scipy.io.wavfile.write(resampled_file, sr, np_buffer.astype(np.int16)) - return resampled_file +import ffmpeg +import whisper_live.utils as utils class Client: @@ -150,10 +96,40 @@ class Client: self.transcript = [] print("[INFO]: * recording") + def handle_status_messages(self, message_data): + """Handles server status messages.""" + status = message_data["status"] + if status == "WAIT": + self.waiting = True + print(f"[INFO]: Server is full. Estimated wait time {round(message_data['message'])} minutes.") + elif status == "ERROR": + print(f"Message from Server: {message_data['message']}") + self.server_error = True + elif status == "WARNING": + print(f"Message from Server: {message_data['message']}") + + def process_segments(self, segments): + """Processes transcript segments.""" + text = [] + for i, seg in enumerate(segments): + if not text or text[-1] != seg["text"]: + text.append(seg["text"]) + if i == len(segments) - 1: + self.last_segment = seg + elif (self.server_backend == "faster_whisper" and + (not self.transcript or + float(seg['start']) >= float(self.transcript[-1]['end']))): + self.transcript.append(seg) + + # Truncate to last 3 entries for brevity. + text = text[-3:] + utils.clear_screen() + utils.print_transcript(text) + def on_message(self, ws, message): """ Callback function called when a message is received from the server. - + It updates various attributes of the client based on the received message, including recording status, language detection, and server messages. If a disconnect message is received, it sets the recording status to False. @@ -171,18 +147,11 @@ class Client: return if "status" in message.keys(): - if message["status"] == "WAIT": - self.waiting = True - print( - f"[INFO]:Server is full. Estimated wait time {round(message['message'])} minutes." - ) - elif message["status"] == "ERROR": - print(f"Message from Server: {message['message']}") - self.server_error = True + self.handle_status_messages(message) return if "message" in message.keys() and message["message"] == "DISCONNECT": - print("[INFO]: Server overtime disconnected.") + print("[INFO]: Server disconnected due to overtime.") self.recording = False if "message" in message.keys() and message["message"] == "SERVER_READY": @@ -199,38 +168,8 @@ class Client: ) return - if "segments" not in message.keys(): - return - - message = message["segments"] - text = [] - n_segments = len(message) - - if n_segments: - for i, seg in enumerate(message): - if text and text[-1] == seg["text"]: - # already got it - continue - text.append(seg["text"]) - - if i == n_segments-1: - self.last_segment = seg - elif self.server_backend == "faster_whisper": - if not len(self.transcript) or float(seg['start']) >= float(self.transcript[-1]['end']): - self.transcript.append(seg) - - # keep only last 3 - if len(text) > 3: - text = text[-3:] - wrapper = textwrap.TextWrapper(width=60) - word_list = wrapper.wrap(text="".join(text)) - # Print each line. - if os.name == "nt": - os.system("cls") - else: - os.system("clear") - for element in word_list: - print(element) + if "segments" in message.keys(): + self.process_segments(message["segments"]) def on_error(self, ws, error): print(f"[ERROR] WebSocket Error: {error}") @@ -246,7 +185,7 @@ class Client: def on_open(self, ws): """ Callback function called when the WebSocket connection is successfully opened. - + Sends an initial configuration message to the server, including client UID, language selection, and task type. @@ -270,8 +209,8 @@ class Client: def bytes_to_float_array(audio_bytes): """ Convert audio data from bytes to a NumPy float array. - - It assumes that the audio data is in 16-bit PCM format. The audio data is normalized to + + It assumes that the audio data is in 16-bit PCM format. The audio data is normalized to have values between -1 and 1. Args: @@ -299,10 +238,10 @@ class Client: def play_file(self, filename): """ Play an audio file and send it to the server for processing. - + Reads an audio file, plays it through the audio output, and simultaneously sends - the audio data to the server for processing. It uses PyAudio to create an audio - stream for playback. The audio data is read from the file in chunks, converted to + the audio data to the server for processing. It uses PyAudio to create an audio + stream for playback. The audio data is read from the file in chunks, converted to floating-point format, and sent to the server using WebSocket communication. This method is typically used when you want to process pre-recorded audio and send it to the server in real-time. @@ -310,7 +249,7 @@ class Client: Args: filename (str): The path to the audio file to be played and sent to the server. """ - + # read audio and create pyaudio stream with wave.open(filename, "rb") as wavfile: self.stream = self.p.open( @@ -356,7 +295,7 @@ class Client: """ Close the WebSocket connection and join the WebSocket thread. - First attempts to close the WebSocket connection using `self.client_socket.close()`. After + First attempts to close the WebSocket connection using `self.client_socket.close()`. After closing the connection, it joins the WebSocket thread to ensure proper termination. """ @@ -383,7 +322,7 @@ class Client: """ Write audio frames to a WAV file. - The WAV file is created or overwritten with the specified name. The audio frames should be + The WAV file is created or overwritten with the specified name. The audio frames should be in the correct format and match the specified channel, sample width, and sample rate. Args: @@ -433,7 +372,6 @@ class Client: print("[INFO]: HLS stream processing finished.") - def record(self, out_file="output_recording.wav"): """ Record audio data from the input stream and save it to a WAV file. @@ -444,11 +382,12 @@ class Client: Audio data is saved in chunks to the "chunks" directory. Each chunk is saved as a separate WAV file. The recording will continue until the specified duration is reached or until the `RECORDING` flag is set to `False`. - The recording process can be interrupted by sending a KeyboardInterrupt (e.g., pressing Ctrl+C). After recording, + The recording process can be interrupted by sending a KeyboardInterrupt (e.g., pressing Ctrl+C). After recording, the method combines all the saved audio chunks into the specified `out_file`. Args: - out_file (str, optional): The name of the output WAV file to save the entire recording. Default is "output_recording.wav". + out_file (str, optional): The name of the output WAV file to save the entire recording. + Default is "output_recording.wav". """ n_audio_file = 0 @@ -458,7 +397,7 @@ class Client: for _ in range(0, int(self.rate / self.chunk * self.record_seconds)): if not self.recording: break - data = self.stream.read(self.chunk, exception_on_overflow = False) + data = self.stream.read(self.chunk, exception_on_overflow=False) self.frames += data audio_array = Client.bytes_to_float_array(data) @@ -498,8 +437,8 @@ class Client: def write_output_recording(self, n_audio_file, out_file): """ Combine and save recorded audio chunks into a single WAV file. - - The individual audio chunk files are expected to be located in the "chunks" directory. Reads each chunk + + The individual audio chunk files are expected to be located in the "chunks" directory. Reads each chunk file, appends its audio data to the final recording, and then deletes the chunk file. After combining and saving, the final recording is stored in the specified `out_file`. @@ -532,7 +471,7 @@ class Client: def write_srt_file(self, output_path="output.srt"): self.transcript.append(self.last_segment) - create_srt_file(self.transcript, output_path) + utils.create_srt_file(self.transcript, output_path) class TranscriptionClient: @@ -558,13 +497,7 @@ class TranscriptionClient: transcription_client() ``` """ - def __init__(self, - host, - port, - lang=None, - translate=False, - model="small", - ): + def __init__(self, host, port, lang=None, translate=False, model="small"): self.client = Client(host, port, lang, translate, model) def __call__(self, audio=None, hls_url=None): @@ -572,12 +505,12 @@ class TranscriptionClient: Start the transcription process. Initiates the transcription process by connecting to the server via a WebSocket. It waits for the server - to be ready to receive audio data and then sends audio for transcription. If an audio file is provided, it + to be ready to receive audio data and then sends audio for transcription. If an audio file is provided, it will be played and streamed to the server; otherwise, it will perform live recording. Args: audio (str, optional): Path to an audio file for transcription. Default is None, which triggers live recording. - + """ print("[INFO]: Waiting for server ready ...") while not self.client.recording: @@ -589,7 +522,7 @@ class TranscriptionClient: if hls_url is not None: self.client.process_hls_stream(hls_url) elif audio is not None: - resampled_file = resample(audio) + resampled_file = utils.resample(audio) self.client.play_file(resampled_file) else: - self.client.record() \ No newline at end of file + self.client.record() diff --git a/whisper_live/server.py b/whisper_live/server.py index e9e8493..dc7d5c3 100644 --- a/whisper_live/server.py +++ b/whisper_live/server.py @@ -2,69 +2,206 @@ import os import time import threading import json -import textwrap - +import functools import logging -logging.basicConfig(level = logging.INFO) - -from websockets.sync.server import serve - import torch import numpy as np - -from whisper_live.vad import VoiceActivityDetection -import functools - -from whisper_live.vad import VoiceActivityDetection +from websockets.sync.server import serve +from websockets.exceptions import ConnectionClosed +from whisper_live.vad import VoiceActivityDetector from whisper_live.transcriber import WhisperModel try: from whisper_live.transcriber_tensorrt import WhisperTRTLLM -except Exception as e: +except Exception: pass +logging.basicConfig(level=logging.INFO) -class TranscriptionServer: - """ - Represents a transcription server that handles incoming audio from clients. - Attributes: - RATE (int): The audio sampling rate (constant) set to 16000. - vad_model (torch.Module): The voice activity detection model. - vad_threshold (float): The voice activity detection threshold. - clients (dict): A dictionary to store connected clients. - websockets (dict): A dictionary to store WebSocket connections. - clients_start_time (dict): A dictionary to track client start times. - max_clients (int): Maximum allowed connected clients. - max_connection_time (int): Maximum allowed connection time in seconds. - """ +class ClientManager: + def __init__(self, max_clients=4, max_connection_time=600): + """ + Initializes the ClientManager with specified limits on client connections and connection durations. - RATE = 16000 - - def __init__(self): - # voice activity detection model - + Args: + max_clients (int, optional): The maximum number of simultaneous client connections allowed. Defaults to 4. + max_connection_time (int, optional): The maximum duration (in seconds) a client can stay connected. Defaults + to 600 seconds (10 minutes). + """ self.clients = {} - self.websockets = {} - self.clients_start_time = {} - self.max_clients = 4 - self.max_connection_time = 600 + self.start_times = {} + self.max_clients = max_clients + self.max_connection_time = max_connection_time + + def add_client(self, websocket, client): + """ + Adds a client and their connection start time to the tracking dictionaries. + + Args: + websocket: The websocket associated with the client to add. + client: The client object to be added and tracked. + """ + self.clients[websocket] = client + self.start_times[websocket] = time.time() + + def get_client(self, websocket): + """ + Retrieves a client associated with the given websocket. + + Args: + websocket: The websocket associated with the client to retrieve. + + Returns: + The client object if found, False otherwise. + """ + if websocket in self.clients: + return self.clients[websocket] + return False + + def remove_client(self, websocket): + """ + Removes a client and their connection start time from the tracking dictionaries. Performs cleanup on the + client if necessary. + + Args: + websocket: The websocket associated with the client to be removed. + """ + client = self.clients.pop(websocket, None) + if client: + client.cleanup() + self.start_times.pop(websocket, None) def get_wait_time(self): """ - Calculate and return the estimated wait time for clients. + Calculates the estimated wait time for new clients based on the remaining connection times of current clients. Returns: - float: The estimated wait time in minutes. + The estimated wait time in minutes for new clients to connect. Returns 0 if there are available slots. """ wait_time = None - - for k, v in self.clients_start_time.items(): - current_client_time_remaining = self.max_connection_time - (time.time() - v) - + for start_time in self.start_times.values(): + current_client_time_remaining = self.max_connection_time - (time.time() - start_time) if wait_time is None or current_client_time_remaining < wait_time: wait_time = current_client_time_remaining + return wait_time / 60 if wait_time is not None else 0 - return wait_time / 60 + def is_server_full(self, websocket, options): + """ + Checks if the server is at its maximum client capacity and sends a wait message to the client if necessary. + + Args: + websocket: The websocket of the client attempting to connect. + options: A dictionary of options that may include the client's unique identifier. + + Returns: + True if the server is full, False otherwise. + """ + if len(self.clients) >= self.max_clients: + wait_time = self.get_wait_time() + response = {"uid": options["uid"], "status": "WAIT", "message": wait_time} + websocket.send(json.dumps(response)) + return True + return False + + def is_client_timeout(self, websocket): + """ + Checks if a client has exceeded the maximum allowed connection time and disconnects them if so, issuing a warning. + + Args: + websocket: The websocket associated with the client to check. + + Returns: + True if the client's connection time has exceeded the maximum limit, False otherwise. + """ + elapsed_time = time.time() - self.start_times[websocket] + if elapsed_time >= self.max_connection_time: + self.clients[websocket].disconnect() + logging.warning(f"Client with uid '{self.clients[websocket].client_uid}' disconnected due to overtime.") + return True + return False + + +class TranscriptionServer: + RATE = 16000 + + def __init__(self): + self.client_manager = ClientManager() + self.no_voice_activity_chunks = 0 + + def initialize_client( + self, websocket, options, faster_whisper_custom_model_path, + whisper_tensorrt_path, trt_multilingual + ): + if self.backend == "tensorrt": + try: + client = ServeClientTensorRT( + websocket, + multilingual=trt_multilingual, + language=options["language"], + task=options["task"], + client_uid=options["uid"], + model=whisper_tensorrt_path + ) + logging.info("Running TensorRT backend.") + except Exception as e: + logging.error(f"TensorRT-LLM not supported: {e}") + self.client_uid = options["uid"] + websocket.send(json.dumps({ + "uid": self.client_uid, + "status": "WARNING", + "message": "TensorRT-LLM not supported on Server yet. " + "Reverting to available backend: 'faster_whisper'" + })) + self.backend = "faster_whisper" + + if self.backend == "faster_whisper": + if faster_whisper_custom_model_path is not None and os.path.exists(faster_whisper_custom_model_path): + logging.info(f"Using custom model {faster_whisper_custom_model_path}") + options["model"] = faster_whisper_custom_model_path + client = ServeClientFasterWhisper( + websocket, + language=options["language"], + task=options["task"], + client_uid=options["uid"], + model=options["model"], + initial_prompt=options.get("initial_prompt"), + vad_parameters=options.get("vad_parameters") + ) + logging.info("Running faster_whisper backend.") + + # self.clients[websocket] = client + # self.clients_start_time[websocket] = time.time() + self.client_manager.add_client(websocket, client) + + def get_audio_from_websocket(self, websocket): + """ + Receives audio buffer from websocket and creates a numpy array out of it. + + Args: + websocket: The websocket to receive audio from. + + Returns: + A numpy array containing the audio. + """ + frame_data = websocket.recv() + return np.frombuffer(frame_data, dtype=np.float32) + + def handle_new_connection(self, websocket, backend, faster_whisper_custom_model_path, + whisper_tensorrt_path, trt_multilingual): + logging.info("New client connected") + options = websocket.recv() + options = json.loads(options) + + if self.client_manager.is_server_full(websocket, options): + websocket.close() + return + + self.backend = backend + if self.backend == "tensorrt": + self.vad_detector = VoiceActivityDetector(frame_rate=self.RATE) + + self.initialize_client( + websocket, options, faster_whisper_custom_model_path, whisper_tensorrt_path, trt_multilingual) def recv_audio(self, websocket, @@ -74,7 +211,7 @@ class TranscriptionServer: trt_multilingual=False): """ Receive audio chunks from a client in an infinite loop. - + Continuously receives audio frames from a connected client over a WebSocket connection. It processes the audio frames using a voice activity detection (VAD) model to determine if they contain speech @@ -96,127 +233,48 @@ class TranscriptionServer: Raises: Exception: If there is an error during the audio frame processing. """ - self.backend = backend - if self.backend == "tensorrt": - self.vad_model = VoiceActivityDetection() - self.vad_threshold = 0.5 + try: + self.handle_new_connection(websocket, backend, faster_whisper_custom_model_path, + whisper_tensorrt_path, trt_multilingual) - logging.info("New client connected") - options = websocket.recv() - options = json.loads(options) + while not self.client_manager.is_client_timeout(websocket): + try: + frame_np = self.get_audio_from_websocket(websocket) + client = self.client_manager.get_client(websocket) - if len(self.clients) >= self.max_clients: - logging.warning("Client Queue Full. Asking client to wait ...") - wait_time = self.get_wait_time() - response = { - "uid": options["uid"], - "status": "WAIT", - "message": wait_time, - } - websocket.send(json.dumps(response)) - websocket.close() - del websocket - return - - if self.backend == "tensorrt": - try: - import tensorrt as trt - import tensorrt_llm - self.backend = "tensorrt" - client = ServeClientTensorRT( - websocket, - multilingual=trt_multilingual, - language=options["language"], - task=options["task"], - client_uid=options["uid"], - model=whisper_tensorrt_path - ) - logging.info(f"Running TensorRT backend.") - except Exception as e: - self.client_uid = options["uid"] - websocket.send( - json.dumps( - { - "uid": self.client_uid, - "status": "ERROR", - "message": f"TensorRT-LLM not supported on Server yet. Reverting to available backend: 'faster_whisper'" - } - ) - ) - self.backend = "faster_whisper" - - if self.backend == "faster_whisper": - # validate custom model - if faster_whisper_custom_model_path is not None and os.path.exists(faster_whisper_custom_model_path): - logging.info(f"Using custom model {faster_whisper_custom_model_path}") - options["model"] = faster_whisper_custom_model_path - client = ServeClientFasterWhisper( - websocket, - language=options["language"], - task=options["task"], - client_uid=options["uid"], - model=options["model"], - initial_prompt=options.get("initial_prompt"), - vad_parameters=options.get("vad_parameters") - ) - logging.info(f"Running faster_whisper backend.") - - self.clients[websocket] = client - self.clients_start_time[websocket] = time.time() - no_voice_activity_chunks = 0 - - while True: - try: - frame_data = websocket.recv() - frame_np = np.frombuffer(frame_data, dtype=np.float32) - - # VAD, for faster_whisper VAD model is already integrated - if self.backend == "tensorrt": - try: - speech_prob = self.vad_model(torch.from_numpy(frame_np.copy()), self.RATE).item() - if speech_prob < self.vad_threshold: - no_voice_activity_chunks += 1 - if no_voice_activity_chunks > 3: - if not self.clients[websocket].eos: - self.clients[websocket].set_eos(True) - time.sleep(0.1) # Sleep 100m; wait some voice activity. + # VAD, for faster_whisper VAD model is already integrated + if self.backend == "tensorrt": + if not self.voice_activity(websocket, frame_np): continue - no_voice_activity_chunks = 0 - self.clients[websocket].set_eos(False) + self.no_voice_activity_chunks = 0 + client.set_eos(False) - except Exception as e: - logging.error(e) - return - - self.clients[websocket].add_frames(frame_np) + client.add_frames(frame_np) - elapsed_time = time.time() - self.clients_start_time[websocket] - if elapsed_time >= self.max_connection_time: - self.clients[websocket].disconnect() - logging.warning(f"Client with uid '{self.clients[websocket].client_uid}' disconnected due to overtime.") - self.clients[websocket].cleanup() - self.clients.pop(websocket) - self.clients_start_time.pop(websocket) + except Exception as e: + logging.error(e) + self.cleanup(websocket) websocket.close() - del websocket break + except ConnectionClosed: + logging.info("Connection closed by client.") + except json.JSONDecodeError: + logging.error("Failed to decode JSON from client") + except Exception as e: + logging.error(f"Unexpected error: {str(e)}") + finally: + if self.client_manager.get_client(websocket): + self.cleanup(websocket) + websocket.close() + del websocket - except Exception as e: - logging.error(e) - self.clients[websocket].cleanup() - self.clients.pop(websocket) - self.clients_start_time.pop(websocket) - del websocket - break - - def run(self, - host, - port=9090, - backend="tensorrt", + def run(self, + host, + port=9090, + backend="tensorrt", faster_whisper_custom_model_path=None, - whisper_tensorrt_path=None, - trt_multilingual=False - ): + whisper_tensorrt_path=None, + trt_multilingual=False): """ Run the transcription server. @@ -237,6 +295,46 @@ class TranscriptionServer: ) as server: server.serve_forever() + def voice_activity(self, websocket, frame_np): + """ + Evaluates the voice activity in a given audio frame and manages the state of voice activity detection. + + This method uses the configured voice activity detection (VAD) model to assess whether the given audio frame + contains speech. If the VAD model detects no voice activity for more than three consecutive frames, + it sets an end-of-speech (EOS) flag for the associated client. This method aims to efficiently manage + speech detection to improve subsequent processing steps. + + Args: + websocket: The websocket associated with the current client. Used to retrieve the client object + from the client manager for state management. + frame_np (numpy.ndarray): The audio frame to be analyzed. This should be a NumPy array containing + the audio data for the current frame. + + Returns: + bool: True if voice activity is detected in the current frame, False otherwise. When returning False + after detecting no voice activity for more than three consecutive frames, it also triggers the + end-of-speech (EOS) flag for the client. + """ + if not self.vad_detector(frame_np): + self.no_voice_activity_chunks += 1 + if self.no_voice_activity_chunks > 3: + client = self.client_manager.get_client(websocket) + if not client.eos: + client.set_eos(True) + time.sleep(0.1) # Sleep 100m; wait some voice activity. + return False + return True + + def cleanup(self, websocket): + """ + Cleans up resources associated with a given client's websocket. + + Args: + websocket: The websocket associated with the client to be cleaned up. + """ + if self.client_manager.get_client(websocket): + self.client_manager.remove_client(websocket) + class ServeClientBase(object): RATE = 16000 @@ -246,7 +344,6 @@ class ServeClientBase(object): def __init__(self, client_uid, websocket): self.client_uid = client_uid self.websocket = websocket - self.data = b"" self.frames = b"" self.timestamp_offset = 0.0 self.frames_np = None @@ -254,7 +351,7 @@ class ServeClientBase(object): self.text = [] self.current_out = '' self.prev_out = '' - self.t_start=None + self.t_start = None self.exit = False self.same_output_threshold = 0 self.show_prev_out_thresh = 5 # if pause(no output from whisper) show previous output for 5 seconds @@ -263,12 +360,20 @@ class ServeClientBase(object): self.send_last_n_segments = 10 # text formatting - self.wrapper = textwrap.TextWrapper(width=50) self.pick_previous_segments = 2 # threading self.lock = threading.Lock() - + + def speech_to_text(self): + raise NotImplementedError + + def transcribe_audio(self): + raise NotImplementedError + + def handle_transcription_output(self): + raise NotImplementedError + def add_frames(self, frame_np): """ Add audio frames to the ongoing audio stream buffer. @@ -295,9 +400,92 @@ class ServeClientBase(object): self.frames_np = np.concatenate((self.frames_np, frame_np), axis=0) self.lock.release() - def speech_to_text(self): - raise NotImplementedError("Please implement in child Class.") - + def clip_audio_if_no_valid_segment(self): + """ + Update the timestamp offset based on audio buffer status. + Clip audio if the current chunk exceeds 30 seconds, this basically implies that + no valid segment for the last 30 seconds from whisper + """ + if self.frames_np[int((self.timestamp_offset - self.frames_offset)*self.RATE):].shape[0] > 25 * self.RATE: + duration = self.frames_np.shape[0] / self.RATE + self.timestamp_offset = self.frames_offset + duration - 5 + + def get_audio_chunk_for_processing(self): + """ + Retrieves the next chunk of audio data for processing based on the current offsets. + + Calculates which part of the audio data should be processed next, based on + the difference between the current timestamp offset and the frame's offset, scaled by + the audio sample rate (RATE). It then returns this chunk of audio data along with its + duration in seconds. + + Returns: + tuple: A tuple containing: + - input_bytes (np.ndarray): The next chunk of audio data to be processed. + - duration (float): The duration of the audio chunk in seconds. + """ + samples_take = max(0, (self.timestamp_offset - self.frames_offset) * self.RATE) + input_bytes = self.frames_np[int(samples_take):].copy() + duration = input_bytes.shape[0] / self.RATE + return input_bytes, duration + + def prepare_segments(self, last_segment=None): + """ + Prepares the segments of transcribed text to be sent to the client. + + This method compiles the recent segments of transcribed text, ensuring that only the + specified number of the most recent segments are included. It also appends the most + recent segment of text if provided (which is considered incomplete because of the possibility + of the last word being truncated in the audio chunk). + + Args: + last_segment (str, optional): The most recent segment of transcribed text to be added + to the list of segments. Defaults to None. + + Returns: + list: A list of transcribed text segments to be sent to the client. + """ + segments = [] + if len(self.transcript) >= self.send_last_n_segments: + segments = self.transcript[-self.send_last_n_segments:].copy() + else: + segments = self.transcript.copy() + if last_segment is not None: + segments = segments + [last_segment] + return segments + + def get_audio_chunk_duration(self, input_bytes): + """ + Calculates the duration of the provided audio chunk. + + Args: + input_bytes (numpy.ndarray): The audio chunk for which to calculate the duration. + + Returns: + float: The duration of the audio chunk in seconds. + """ + return input_bytes.shape[0] / self.RATE + + def send_transcription_to_client(self, segments): + """ + Sends the specified transcription segments to the client over the websocket connection. + + This method formats the transcription segments into a JSON object and attempts to send + this object to the client. If an error occurs during the send operation, it logs the error. + + Returns: + segments (list): A list of transcription segments to be sent to the client. + """ + try: + self.websocket.send( + json.dumps({ + "uid": self.client_uid, + "segments": segments, + }) + ) + except Exception as e: + logging.error(f"[ERROR]: Sending data to client: {e}") + def disconnect(self): """ Notify the client of disconnection and send a disconnect message. @@ -306,15 +494,11 @@ class ServeClientBase(object): that the transcription service is disconnecting gracefully. """ - self.websocket.send( - json.dumps( - { - "uid": self.client_uid, - "message": self.DISCONNECT - } - ) - ) - + self.websocket.send(json.dumps({ + "uid": self.client_uid, + "message": self.DISCONNECT + })) + def cleanup(self): """ Perform cleanup tasks before exiting the transcription service. @@ -329,44 +513,7 @@ class ServeClientBase(object): class ServeClientTensorRT(ServeClientBase): - """ - Attributes: - RATE (int): The audio sampling rate (constant) set to 16000. - SERVER_READY (str): A constant message indicating that the server is ready. - DISCONNECT (str): A constant message indicating that the client should disconnect. - client_uid (str): A unique identifier for the client. - data (bytes): Accumulated audio data. - frames (bytes): Accumulated audio frames. - language (str): The language for transcription. - task (str): The task type, e.g., "transcribe." - transcriber (WhisperModel): The Whisper model for speech-to-text. - timestamp_offset (float): The offset in audio timestamps. - frames_np (numpy.ndarray): NumPy array to store audio frames. - frames_offset (float): The offset in audio frames. - text (list): List of transcribed text segments. - current_out (str): The current incomplete transcription. - prev_out (str): The previous incomplete transcription. - t_start (float): Timestamp for the start of transcription. - exit (bool): A flag to exit the transcription thread. - same_output_threshold (int): Threshold for consecutive same output segments. - show_prev_out_thresh (int): Threshold for showing previous output segments. - add_pause_thresh (int): Threshold for adding a pause (blank) segment. - transcript (list): List of transcribed segments. - send_last_n_segments (int): Number of last segments to send to the client. - wrapper (textwrap.TextWrapper): Text wrapper for formatting text. - pick_previous_segments (int): Number of previous segments to include in the output. - websocket: The WebSocket connection for the client. - """ - def __init__( - self, - websocket, - task="transcribe", - device=None, - multilingual=False, - language=None, - client_uid=None, - model=None - ): + def __init__(self, websocket, task="transcribe", multilingual=False, language=None, client_uid=None, model=None): """ Initialize a ServeClient instance. The Whisper model is initialized based on the client's language and device availability. @@ -387,8 +534,8 @@ class ServeClientTensorRT(ServeClientBase): self.task = task self.eos = False self.transcriber = WhisperTRTLLM( - model, - assets_dir="assets", + model, + assets_dir="assets", device="cuda", is_multilingual=multilingual, language=self.language, @@ -400,52 +547,75 @@ class ServeClientTensorRT(ServeClientBase): self.trans_thread = threading.Thread(target=self.speech_to_text) self.trans_thread.start() - self.websocket.send( - json.dumps( - { - "uid": self.client_uid, - "message": self.SERVER_READY, - "backend": "tensorrt" - } - ) - ) + self.websocket.send(json.dumps({ + "uid": self.client_uid, + "message": self.SERVER_READY, + "backend": "tensorrt" + })) def warmup(self, warmup_steps=10): + """ + Warmup TensorRT since first few inferences are slow. + + Args: + warmup_steps (int): Number of steps to warm up the model for. + """ logging.info("[INFO:] Warming up TensorRT engine..") mel, _ = self.transcriber.log_mel_spectrogram("tests/jfk.flac") for i in range(warmup_steps): self.transcriber.transcribe(mel) - + def set_eos(self, eos): + """ + Sets the End of Speech (EOS) flag. + + Args: + eos (bool): The value to set for the EOS flag. + """ self.lock.acquire() self.eos = eos self.lock.release() - - def add_frames(self, frame_np): + + def handle_transcription_output(self, last_segment, duration): """ - Add audio frames to the ongoing audio stream buffer. - - This method is responsible for maintaining the audio stream buffer, allowing the continuous addition - of audio frames as they are received. It also ensures that the buffer does not exceed a specified size - to prevent excessive memory usage. - - If the buffer size exceeds a threshold (45 seconds of audio data), it discards the oldest 30 seconds - of audio data to maintain a reasonable buffer size. If the buffer is empty, it initializes it with the provided - audio frame. The audio stream buffer is used for real-time processing of audio data for transcription. + Handle the transcription output, updating the transcript and sending data to the client. Args: - frame_np (numpy.ndarray): The audio frame data as a NumPy array. - + last_segment (str): The last segment from the whisper output which is considered to be incomplete because + of the possibility of word being truncated. + duration (float): Duration of the transcribed audio chunk. """ - self.lock.acquire() - if self.frames_np is not None and self.frames_np.shape[0] > 45*self.RATE: - self.frames_offset += 30.0 - self.frames_np = self.frames_np[int(30*self.RATE):] - if self.frames_np is None: - self.frames_np = frame_np.copy() - else: - self.frames_np = np.concatenate((self.frames_np, frame_np), axis=0) - self.lock.release() + segments = self.prepare_segments({"text": last_segment}) + self.send_transcription_to_client(segments) + if self.eos: + self.update_timestamp_offset(last_segment, duration) + + def transcribe_audio(self, input_bytes): + """ + Transcribe the audio chunk and send the results to the client. + + Args: + input_bytes (np.array): The audio chunk to transcribe. + """ + logging.info(f"[WhisperTensorRT:] Processing audio with duration: {input_bytes.shape[0] / self.RATE}") + mel, duration = self.transcriber.log_mel_spectrogram(input_bytes) + last_segment = self.transcriber.transcribe(mel) + if last_segment: + self.handle_transcription_output(last_segment, duration) + + def update_timestamp_offset(self, last_segment, duration): + """ + Update timestamp offset and transcript. + + Args: + last_segment (str): Last transcribed audio from the whisper model. + duration (float): Duration of the last audio chunk. + """ + if not len(self.transcript): + self.transcript.append({"text": last_segment + " "}) + elif self.transcript[-1]["text"].strip() != last_segment: + self.transcript.append({"text": last_segment + " "}) + self.timestamp_offset += duration def speech_to_text(self): """ @@ -456,8 +626,8 @@ class ServeClientTensorRT(ServeClientBase): If the client's language is not detected, it waits for 30 seconds of audio input to make a language prediction. It utilizes the Whisper ASR model to transcribe the audio, continuously processing and streaming results. Segments - are sent to the client in real-time, and a history of segments is maintained to provide context.Pauses in speech - (no output from Whisper) are handled by showing the previous output for a set duration. A blank segment is added if + are sent to the client in real-time, and a history of segments is maintained to provide context.Pauses in speech + (no output from Whisper) are handled by showing the previous output for a set duration. A blank segment is added if there is no speech for a specified duration to indicate a pause. Raises: @@ -468,99 +638,29 @@ class ServeClientTensorRT(ServeClientBase): if self.exit: logging.info("Exiting speech to text thread") break - + if self.frames_np is None: time.sleep(0.02) # wait for any audio to arrive continue - # clip audio if the current chunk exceeds 30 seconds, this basically implies that - # no valid segment for the last 30 seconds from whisper - if self.frames_np[int((self.timestamp_offset - self.frames_offset)*self.RATE):].shape[0] > 25 * self.RATE: - duration = self.frames_np.shape[0] / self.RATE - self.timestamp_offset = self.frames_offset + duration - 5 - - samples_take = max(0, (self.timestamp_offset - self.frames_offset)*self.RATE) - input_bytes = self.frames_np[int(samples_take):].copy() - duration = input_bytes.shape[0] / self.RATE - if duration<0.4: + self.clip_audio_if_no_valid_segment() + + input_bytes, duration = self.get_audio_chunk_for_processing() + if duration < 0.4: continue try: input_sample = input_bytes.copy() logging.info(f"[WhisperTensorRT:] Processing audio with duration: {duration}") - mel, duration = self.transcriber.log_mel_spectrogram(input_sample) - last_segment = self.transcriber.transcribe(mel) - segments = [] - if len(last_segment): - if len(self.transcript) < self.send_last_n_segments: - segments = self.transcript[:].copy() - else: - segments = self.transcript[-self.send_last_n_segments:].copy() - if last_segment is not None: - segments.append({"text": last_segment}) - try: - self.websocket.send( - json.dumps({ - "uid": self.client_uid, - "segments": segments, - }) - ) - - if self.eos: - if not len(self.transcript): - self.transcript.append({"text": last_segment + " "}) - elif self.transcript[-1]["text"].strip() != last_segment: - self.transcript.append({"text": last_segment + " "}) - self.timestamp_offset += duration - - - except Exception as e: - logging.error(f"[ERROR]: {e}") + self.transcribe_audio(input_sample) except Exception as e: logging.error(f"[ERROR]: {e}") class ServeClientFasterWhisper(ServeClientBase): - """ - Attributes: - RATE (int): The audio sampling rate (constant) set to 16000. - SERVER_READY (str): A constant message indicating that the server is ready. - DISCONNECT (str): A constant message indicating that the client should disconnect. - client_uid (str): A unique identifier for the client. - data (bytes): Accumulated audio data. - frames (bytes): Accumulated audio frames. - language (str): The language for transcription. - task (str): The task type, e.g., "transcribe." - transcriber (WhisperModel): The Whisper model for speech-to-text. - timestamp_offset (float): The offset in audio timestamps. - frames_np (numpy.ndarray): NumPy array to store audio frames. - frames_offset (float): The offset in audio frames. - text (list): List of transcribed text segments. - current_out (str): The current incomplete transcription. - prev_out (str): The previous incomplete transcription. - t_start (float): Timestamp for the start of transcription. - exit (bool): A flag to exit the transcription thread. - same_output_threshold (int): Threshold for consecutive same output segments. - show_prev_out_thresh (int): Threshold for showing previous output segments. - add_pause_thresh (int): Threshold for adding a pause (blank) segment. - transcript (list): List of transcribed segments. - send_last_n_segments (int): Number of last segments to send to the client. - wrapper (textwrap.TextWrapper): Text wrapper for formatting text. - pick_previous_segments (int): Number of previous segments to include in the output. - websocket: The WebSocket connection for the client. - """ - def __init__( - self, - websocket, - task="transcribe", - device=None, - language=None, - client_uid=None, - model="small.en", - initial_prompt=None, - vad_parameters=None, - ): + def __init__(self, websocket, task="transcribe", device=None, language=None, client_uid=None, model="small.en", + initial_prompt=None, vad_parameters=None): """ Initialize a ServeClient instance. The Whisper model is initialized based on the client's language and device availability. @@ -573,7 +673,8 @@ class ServeClientFasterWhisper(ServeClientBase): device (str, optional): The device type for Whisper, "cuda" or "cpu". Defaults to None. language (str, optional): The language for transcription. Defaults to None. client_uid (str, optional): A unique identifier for the client. Defaults to None. - + model (str, optional): The whisper model size. Defaults to 'small.en' + initial_prompt (str, optional): Prompt for whisper inference. Defaults to None. """ super().__init__(client_uid, websocket) self.model_sizes = [ @@ -589,16 +690,16 @@ class ServeClientFasterWhisper(ServeClientBase): self.initial_prompt = initial_prompt self.vad_parameters = vad_parameters or {"threshold": 0.5} self.no_speech_thresh = 0.45 - + device = "cuda" if torch.cuda.is_available() else "cpu" - - if self.model_size_or_path == None: + + if self.model_size_or_path is None: return self.transcriber = WhisperModel( - self.model_size_or_path, + self.model_size_or_path, device=device, - compute_type="int8" if device=="cpu" else "float16", + compute_type="int8" if device == "cpu" else "float16", local_files_only=False, ) @@ -614,7 +715,7 @@ class ServeClientFasterWhisper(ServeClientBase): } ) ) - + def check_valid_model(self, model_size): """ Check if it's a valid whisper model size. @@ -637,7 +738,97 @@ class ServeClientFasterWhisper(ServeClientBase): ) return None return model_size - + + def set_language(self, info): + """ + Updates the language attribute based on the detected language information. + + Args: + info (object): An object containing the detected language and its probability. This object + must have at least two attributes: `language`, a string indicating the detected + language, and `language_probability`, a float representing the confidence level + of the language detection. + """ + if info.language_probability > 0.5: + self.language = info.language + logging.info(f"Detected language {self.language} with probability {info.language_probability}") + self.websocket.send(json.dumps( + {"uid": self.client_uid, "language": self.language, "language_prob": info.language_probability})) + + def transcribe_audio(self, input_sample): + """ + Transcribes the provided audio sample using the configured transcriber instance. + + If the language has not been set, it updates the session's language based on the transcription + information. + + Args: + input_sample (np.array): The audio chunk to be transcribed. This should be a NumPy + array representing the audio data. + + Returns: + The transcription result from the transcriber. The exact format of this result + depends on the implementation of the `transcriber.transcribe` method but typically + includes the transcribed text. + """ + result, info = self.transcriber.transcribe( + input_sample, + initial_prompt=self.initial_prompt, + language=self.language, + task=self.task, + vad_filter=True, + vad_parameters=self.vad_parameters) + if self.language is None: + self.set_language(info) + return result + + def get_previous_output(self): + """ + Retrieves previously generated transcription outputs if no new transcription is available + from the current audio chunks. + + Checks the time since the last transcription output and, if it is within a specified + threshold, returns the most recent segments of transcribed text. It also manages + adding a pause (blank segment) to indicate a significant gap in speech based on a defined + threshold. + + Returns: + segments (list): A list of transcription segments. This may include the most recent + transcribed text segments or a blank segment to indicate a pause + in speech. + """ + segments = [] + if self.t_start is None: + self.t_start = time.time() + if time.time() - self.t_start < self.show_prev_out_thresh: + segments = self.prepare_segments() + + # add a blank if there is no speech for 3 seconds + if len(self.text) and self.text[-1] != '': + if time.time() - self.t_start > self.add_pause_thresh: + self.text.append('') + return segments + + def handle_transcription_output(self, result, duration): + """ + Handle the transcription output, updating the transcript and sending data to the client. + + Args: + result (str): The result from whisper inference i.e. the list of segments. + duration (float): Duration of the transcribed audio chunk. + """ + segments = [] + if len(result): + self.t_start = None + last_segment = self.update_segments(result, duration) + segments = self.prepare_segments(last_segment) + else: + # show previous output if there is pause i.e. no output from whisper + segments = self.get_previous_output() + + if len(segments): + self.send_transcription_to_client(segments) + def speech_to_text(self): """ Process an audio stream in an infinite loop, continuously transcribing the speech. @@ -647,8 +838,8 @@ class ServeClientFasterWhisper(ServeClientBase): If the client's language is not detected, it waits for 30 seconds of audio input to make a language prediction. It utilizes the Whisper ASR model to transcribe the audio, continuously processing and streaming results. Segments - are sent to the client in real-time, and a history of segments is maintained to provide context.Pauses in speech - (no output from Whisper) are handled by showing the previous output for a set duration. A blank segment is added if + are sent to the client in real-time, and a history of segments is maintained to provide context.Pauses in speech + (no output from Whisper) are handled by showing the previous output for a set duration. A blank segment is added if there is no speech for a specified duration to indicate a pause. Raises: @@ -659,85 +850,41 @@ class ServeClientFasterWhisper(ServeClientBase): if self.exit: logging.info("Exiting speech to text thread") break - - if self.frames_np is None: + + if self.frames_np is None: continue - # clip audio if the current chunk exceeds 30 seconds, this basically implies that - # no valid segment for the last 30 seconds from whisper - if self.frames_np[int((self.timestamp_offset - self.frames_offset)*self.RATE):].shape[0] > 25 * self.RATE: - duration = self.frames_np.shape[0] / self.RATE - self.timestamp_offset = self.frames_offset + duration - 5 - - samples_take = max(0, (self.timestamp_offset - self.frames_offset)*self.RATE) - input_bytes = self.frames_np[int(samples_take):].copy() - duration = input_bytes.shape[0] / self.RATE - if duration<1.0: + self.clip_audio_if_no_valid_segment() + + input_bytes, duration = self.get_audio_chunk_for_processing() + if duration < 1.0: continue try: input_sample = input_bytes.copy() - - # whisper transcribe with prompt - result, info = self.transcriber.transcribe( - input_sample, - initial_prompt=self.initial_prompt, - language=self.language, - task=self.task, - vad_filter=True, - vad_parameters=self.vad_parameters - ) + result = self.transcribe_audio(input_sample) if self.language is None: - if info.language_probability > 0.5: - self.language = info.language - logging.info(f"Detected language {self.language} with probability {info.language_probability}") - self.websocket.send(json.dumps( - {"uid": self.client_uid, "language": self.language, "language_prob": info.language_probability})) - else: - # detect language again - continue - - if len(result): - self.t_start = None - last_segment = self.update_segments(result, duration) - if len(self.transcript) < self.send_last_n_segments: - segments = self.transcript - else: - segments = self.transcript[-self.send_last_n_segments:] - if last_segment is not None: - segments = segments + [last_segment] - else: - # show previous output if there is pause i.e. no output from whisper - segments = [] - if self.t_start is None: self.t_start = time.time() - if time.time() - self.t_start < self.show_prev_out_thresh: - if len(self.transcript) < self.send_last_n_segments: - segments = self.transcript - else: - segments = self.transcript[-self.send_last_n_segments:] - - # add a blank if there is no speech for 3 seconds - if len(self.text) and self.text[-1] != '': - if time.time() - self.t_start > self.add_pause_thresh: - self.text.append('') - - if not len(segments): continue - try: - self.websocket.send( - json.dumps({ - "uid": self.client_uid, - "segments": segments - }) - ) - except Exception as e: - logging.error(f"[ERROR]: Failed to send message to client: {e}") + continue + self.handle_transcription_output(result, duration) except Exception as e: logging.error(f"[ERROR]: Failed to transcribe audio chunk: {e}") time.sleep(0.01) - + def format_segment(self, start, end, text): - """Helper function to format a segment with string timestamps.""" + """ + Formats a transcription segment with precise start and end times alongside the transcribed text. + + Args: + start (float): The start time of the transcription segment in seconds. + end (float): The end time of the transcription segment in seconds. + text (str): The transcribed text corresponding to the segment. + + Returns: + dict: A dictionary representing the formatted transcription segment, including + 'start' and 'end' times as strings with three decimal places and the 'text' + of the transcription. + """ return { 'start': "{:.3f}".format(start), 'end': "{:.3f}".format(end), @@ -750,17 +897,17 @@ class ServeClientFasterWhisper(ServeClientBase): except for the last segment assuming that it is incomplete. Updates the ongoing transcript with transcribed segments, including their start and end times. - Complete segments are appended to the transcript in chronological order. Incomplete segments - (assumed to be the last one) are processed to identify repeated content. If the same incomplete + Complete segments are appended to the transcript in chronological order. Incomplete segments + (assumed to be the last one) are processed to identify repeated content. If the same incomplete segment is seen multiple times, it updates the offset and appends the segment to the transcript. A threshold is used to detect repeated content and ensure it is only included once in the transcript. - The timestamp offset is updated based on the duration of processed segments. The method returns the + The timestamp offset is updated based on the duration of processed segments. The method returns the last processed segment, allowing it to be sent to the client for real-time updates. Args: segments(dict) : dictionary of segments as returned by whisper duration(float): duration of the current chunk - + Returns: dict or None: The last processed segment with its start time, end time, and transcribed text. Returns None if there are no valid segments to process. @@ -775,11 +922,12 @@ class ServeClientFasterWhisper(ServeClientBase): self.text.append(text_) start, end = self.timestamp_offset + s.start, self.timestamp_offset + min(duration, s.end) - if start >= end: continue - if s.no_speech_prob > self.no_speech_thresh: continue + if start >= end: + continue + if s.no_speech_prob > self.no_speech_thresh: + continue self.transcript.append(self.format_segment(start, end, text_)) - offset = min(duration, s.end) self.current_out += segments[-1].text @@ -788,16 +936,16 @@ class ServeClientFasterWhisper(ServeClientBase): self.timestamp_offset + min(duration, segments[-1].end), self.current_out ) - + # if same incomplete segment is seen multiple times then update the offset # and append the segment to the list - if self.current_out.strip() == self.prev_out.strip() and self.current_out != '': + if self.current_out.strip() == self.prev_out.strip() and self.current_out != '': self.same_output_threshold += 1 - else: + else: self.same_output_threshold = 0 - + if self.same_output_threshold > 5: - if not len(self.text) or self.text[-1].strip().lower()!=self.current_out.strip().lower(): + if not len(self.text) or self.text[-1].strip().lower() != self.current_out.strip().lower(): self.text.append(self.current_out) self.transcript.append(self.format_segment( self.timestamp_offset, @@ -810,7 +958,7 @@ class ServeClientFasterWhisper(ServeClientBase): last_segment = None else: self.prev_out = self.current_out - + # update offset if offset is not None: self.timestamp_offset += offset diff --git a/whisper_live/tensorrt_utils.py b/whisper_live/tensorrt_utils.py index 7b21010..9752e7a 100644 --- a/whisper_live/tensorrt_utils.py +++ b/whisper_live/tensorrt_utils.py @@ -214,7 +214,7 @@ def store_transcripts(filename: Pathlike, texts: Iterable[Tuple[str, str, print(f"{cut_id}:\thyp={hyp}", file=f) -def write_error_stats( +def write_error_stats( # noqa: C901 f: TextIO, test_set_name: str, results: List[Tuple[str, str]], @@ -362,4 +362,4 @@ def write_error_stats( hyp_count = corr + hyp_sub + ins print(f"{word} {corr} {tot_errs} {ref_count} {hyp_count}", file=f) - return float(tot_err_rate) \ No newline at end of file + return float(tot_err_rate) diff --git a/whisper_live/transcriber.py b/whisper_live/transcriber.py index e6f2f1c..bd2e082 100644 --- a/whisper_live/transcriber.py +++ b/whisper_live/transcriber.py @@ -400,7 +400,7 @@ class WhisperModel: return segments, info - def generate_segments( + def generate_segments( # noqa: C901 self, features: np.ndarray, tokenizer: Tokenizer, @@ -425,7 +425,7 @@ class WhisperModel: all_segments = [] while seek < content_frames: time_offset = seek * self.feature_extractor.time_per_frame - segment = features[:, seek : seek + self.feature_extractor.nb_max_frames] + segment = features[:, seek:seek + self.feature_extractor.nb_max_frames] segment_size = min( self.feature_extractor.nb_max_frames, content_frames - seek ) @@ -749,7 +749,7 @@ class WhisperModel: if previous_tokens: prompt.append(tokenizer.sot_prev) - prompt.extend(previous_tokens[-(self.max_length // 2 - 1) :]) + prompt.extend(previous_tokens[-(self.max_length // 2 - 1):]) prompt.extend(tokenizer.sot_sequence) @@ -766,7 +766,7 @@ class WhisperModel: return prompt - def add_word_timestamps( + def add_word_timestamps( # noqa: C901 self, segments: List[dict], tokenizer: Tokenizer, diff --git a/whisper_live/transcriber_tensorrt.py b/whisper_live/transcriber_tensorrt.py index a36bb72..aaa8cc1 100644 --- a/whisper_live/transcriber_tensorrt.py +++ b/whisper_live/transcriber_tensorrt.py @@ -1,17 +1,14 @@ -import argparse import json import re -import time from collections import OrderedDict from pathlib import Path -from typing import Dict, Iterable, List, Optional, TextIO, Tuple, Union +from typing import Union import torch import numpy as np +import torch.nn.functional as F from whisper.tokenizer import get_tokenizer -from whisper_live.tensorrt_utils import (mel_filters, store_transcripts, - write_error_stats, load_audio_wav_format, - pad_or_trim, load_audio) +from whisper_live.tensorrt_utils import (mel_filters, load_audio_wav_format, pad_or_trim, load_audio) import tensorrt_llm import tensorrt_llm.logger as logger @@ -38,8 +35,6 @@ class WhisperEncoding: with open(config_path, 'r') as f: config = json.load(f) - use_gpt_attention_plugin = config['plugin_config'][ - 'gpt_attention_plugin'] dtype = config['builder_config']['precision'] n_mels = config['builder_config']['n_mels'] num_languages = config['builder_config']['num_languages'] @@ -176,16 +171,8 @@ class WhisperDecoding: class WhisperTRTLLM(object): - def __init__( - self, - engine_dir, - debug_mode=False, - assets_dir=None, - device=None, - is_multilingual=False, - language="en", - task="transcribe" - ): + def __init__(self, engine_dir, assets_dir=None, device=None, is_multilingual=False, + language="en", task="transcribe"): world_size = 1 runtime_rank = tensorrt_llm.mpi_rank() runtime_mapping = tensorrt_llm.Mapping(world_size, runtime_rank) @@ -212,7 +199,7 @@ class WhisperTRTLLM(object): self, audio: Union[str, np.ndarray, torch.Tensor], padding: int = 0, - return_duration = True + return_duration=True ): """ Compute the log-Mel spectrogram of @@ -242,8 +229,7 @@ class WhisperTRTLLM(object): audio, _ = load_audio_wav_format(audio) else: audio = load_audio(audio) - assert isinstance(audio, - np.ndarray), f"Unsupported audio type: {type(audio)}" + assert isinstance(audio, np.ndarray), f"Unsupported audio type: {type(audio)}" duration = audio.shape[-1] / SAMPLE_RATE audio = pad_or_trim(audio, N_SAMPLES) audio = audio.astype(np.float32) @@ -254,14 +240,9 @@ class WhisperTRTLLM(object): if padding > 0: audio = F.pad(audio, (0, padding)) window = torch.hann_window(N_FFT).to(audio.device) - stft = torch.stft(audio, - N_FFT, - HOP_LENGTH, - window=window, - return_complex=True) + stft = torch.stft(audio, N_FFT, HOP_LENGTH, window=window, return_complex=True) magnitudes = stft[..., :-1].abs()**2 - mel_spec = self.filters @ magnitudes log_spec = torch.clamp(mel_spec, min=1e-10).log10() @@ -272,7 +253,6 @@ class WhisperTRTLLM(object): else: return log_spec - def process_batch( self, mel, @@ -296,7 +276,7 @@ class WhisperTRTLLM(object): text = self.tokenizer.decode(output_ids[i][0]).strip() texts.append(text) return texts - + def transcribe( self, mel, @@ -336,5 +316,5 @@ def decode_wav_file( prediction = re.sub(r'<\|.*?\|>', '', prediction) if normalizer: prediction = normalizer(prediction) - + return prediction.strip() diff --git a/whisper_live/utils.py b/whisper_live/utils.py new file mode 100644 index 0000000..d32105a --- /dev/null +++ b/whisper_live/utils.py @@ -0,0 +1,71 @@ +import os +import textwrap +import scipy +import ffmpeg +import numpy as np + + +def clear_screen(): + """Clears the console screen.""" + os.system("cls" if os.name == "nt" else "clear") + + +def print_transcript(text): + """Prints formatted transcript text.""" + wrapper = textwrap.TextWrapper(width=60) + for line in wrapper.wrap(text="".join(text)): + print(line) + + +def format_time(s): + """Convert seconds (float) to SRT time format.""" + hours = int(s // 3600) + minutes = int((s % 3600) // 60) + seconds = int(s % 60) + milliseconds = int((s - int(s)) * 1000) + return f"{hours:02}:{minutes:02}:{seconds:02},{milliseconds:03}" + + +def create_srt_file(segments, output_file): + with open(output_file, 'w', encoding='utf-8') as srt_file: + segment_number = 1 + for segment in segments: + start_time = format_time(float(segment['start'])) + end_time = format_time(float(segment['end'])) + text = segment['text'] + + srt_file.write(f"{segment_number}\n") + srt_file.write(f"{start_time} --> {end_time}\n") + srt_file.write(f"{text}\n\n") + + segment_number += 1 + + +def resample(file: str, sr: int = 16000): + """ + # https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/audio.py#L22 + Open an audio file and read as mono waveform, resampling as necessary, + save the resampled audio + + Args: + file (str): The audio file to open + sr (int): The sample rate to resample the audio if necessary + + Returns: + resampled_file (str): The resampled audio file + """ + try: + # This launches a subprocess to decode audio while down-mixing and resampling as necessary. + # Requires the ffmpeg CLI and `ffmpeg-python` package to be installed. + out, _ = ( + ffmpeg.input(file, threads=0) + .output("-", format="s16le", acodec="pcm_s16le", ac=1, ar=sr) + .run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True) + ) + except ffmpeg.Error as e: + raise RuntimeError(f"Failed to load audio: {e.stderr.decode()}") from e + np_buffer = np.frombuffer(out, dtype=np.int16) + + resampled_file = f"{file.split('.')[0]}_resampled.wav" + scipy.io.wavfile.write(resampled_file, sr, np_buffer.astype(np.int16)) + return resampled_file diff --git a/whisper_live/vad.py b/whisper_live/vad.py index 31edac8..94e7373 100644 --- a/whisper_live/vad.py +++ b/whisper_live/vad.py @@ -34,7 +34,7 @@ class VoiceActivityDetection(): if sr != 16000 and (sr % 16000 == 0): step = sr // 16000 - x = x[:,::step] + x = x[:, ::step] sr = 16000 if sr not in self.sample_rates: @@ -110,4 +110,33 @@ class VoiceActivityDetection(): subprocess.run(["wget", "-O", model_filename, model_url], check=True) except subprocess.CalledProcessError: print("Failed to download the model using wget.") - return model_filename \ No newline at end of file + return model_filename + + +class VoiceActivityDetector: + def __init__(self, threshold=0.5, frame_rate=16000): + """ + Initializes the VoiceActivityDetector with a voice activity detection model and a threshold. + + Args: + threshold (float, optional): The probability threshold for detecting voice activity. Defaults to 0.5. + """ + self.model = VoiceActivityDetection() + self.threshold = threshold + self.frame_rate = frame_rate + + def __call__(self, audio_frame): + """ + Determines if the given audio frame contains speech by comparing the detected speech probability against + the threshold. + + Args: + audio_frame (np.ndarray): The audio frame to be analyzed for voice activity. It is expected to be a + NumPy array of audio samples. + + Returns: + bool: True if the speech probability exceeds the threshold, indicating the presence of voice activity; + False otherwise. + """ + speech_prob = self.model(torch.from_numpy(audio_frame), self.frame_rate).item() + return speech_prob > self.threshold