cjones/WhisperLive
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'upstream/main' into vad_warnings

Browse Source
This commit is contained in:
makaveli10
2023-09-27 20:35:38 +08:00
parent 5fb6be41bb 1070f7785e
commit 6a087fa1f6
26 changed files with 4158 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files
whisper_live/client.py
+169 -13
View File
@@ -18,12 +18,13 @@ 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
Parameters
----------
file: str
The audio file to open
sr: int
The sample rate to resample the audio if necessary
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.
@@ -43,11 +44,28 @@ def resample(file: str, sr: int = 16000):
class Client:
"""
Handles audio recording, streaming, and communication with a server using WebSocket.
"""
INSTANCES = {}
def __init__(
self, host=None, port=None, is_multilingual=False, lang=None, translate=False
):
"""
Initializes a Client instance for audio recording and streaming to a server.
If host and port are not provided, the WebSocket connection will not be established.
When translate is True, the task will be set to "translate" instead of "transcribe".
he audio recording starts immediately upon initialization.
Args:
host (str): The hostname or IP address of the server.
port (int): The port number for the WebSocket server.
is_multilingual (bool, optional): Specifies if multilingual transcription is enabled. Default is False.
lang (str, optional): The selected language for transcription when multilingual is disabled. Default is None.
translate (bool, optional): Specifies if the task is translation. Default is False.
"""
self.chunk = 1024
self.format = pyaudio.paInt16
self.channels = 1
@@ -77,7 +95,6 @@ class Client:
frames_per_buffer=self.chunk,
)
# create websocket connection
if host is not None and port is not None:
socket_url = f"ws://{host}:{port}"
self.client_socket = websocket.WebSocketApp(
@@ -104,6 +121,18 @@ class Client:
print("[INFO]: * recording")
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.
Args:
ws (websocket.WebSocketApp): The WebSocket client instance.
message (str): The received message from the server.
"""
self.last_response_recieved = time.time()
message = json.loads(message)
@@ -164,6 +193,16 @@ class Client:
print(f"[INFO]: Websocket connection closed: {close_status_code}: {close_msg}")
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, multilingual mode,
language selection, and task type.
Args:
ws (websocket.WebSocketApp): The WebSocket client instance.
"""
print(self.multilingual, self.language, self.task)
print("[INFO]: Opened connection")
@@ -180,16 +219,49 @@ class Client:
@staticmethod
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
have values between -1 and 1.
Args:
audio_bytes (bytes): Audio data in bytes.
Returns:
np.ndarray: A NumPy array containing the audio data as float values normalized between -1 and 1.
"""
raw_data = np.frombuffer(buffer=audio_bytes, dtype=np.int16)
return raw_data.astype(np.float32) / 32768.0
def send_packet_to_server(self, message):
"""
Send an audio packet to the server using WebSocket.
Args:
message (bytes): The audio data packet in bytes to be sent to the server.
"""
try:
self.client_socket.send(message, websocket.ABNF.OPCODE_BINARY)
except Exception as e:
print(e)
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
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.
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(
@@ -213,8 +285,7 @@ class Client:
wavfile.close()
assert self.last_response_recieved
elapsed_time = time.time() - self.last_response_recieved
while elapsed_time < self.disconnect_if_no_response_for:
while time.time() - self.last_response_recieved < self.disconnect_if_no_response_for:
continue
self.stream.close()
self.close_websocket()
@@ -228,20 +299,44 @@ class Client:
print("[INFO]: Keyboard interrupt.")
def close_websocket(self):
"""
Close the WebSocket connection and join the WebSocket thread.
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.
"""
try:
self.client_socket.close() # Close the WebSocket connection
self.client_socket.close()
except Exception as e:
print("[ERROR]: Error closing WebSocket:", e)
try:
self.ws_thread.join() # Wait for the WebSocket thread to finish
self.ws_thread.join()
except Exception as e:
print("[ERROR:] Error joining WebSocket thread:", e)
def get_client_socket(self):
"""
Get the WebSocket client socket instance.
Returns:
WebSocketApp: The WebSocket client socket instance currently in use by the client.
"""
return self.client_socket
def write_audio_frames_to_file(self, frames, file_name):
"""
Write audio frames to a WAV file.
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:
frames (bytes): The audio frames to be written to the file.
file_name (str): The name of the WAV file to which the frames will be written.
"""
with wave.open(file_name, "wb") as wavfile:
wavfile: wave.Wave_write
wavfile.setnchannels(self.channels)
@@ -250,8 +345,23 @@ class Client:
wavfile.writeframes(frames)
def record(self, out_file="output_recording.wav"):
"""
Record audio data from the input stream and save it to a WAV file.
Continuously records audio data from the input stream, sends it to the server via a WebSocket
connection, and simultaneously saves it to multiple WAV files in chunks. It stops recording when
the `RECORD_SECONDS` duration is reached or when the `RECORDING` flag is set to `False`.
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 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".
"""
n_audio_file = 0
# create dir for saving audio chunks
if not os.path.exists("chunks"):
os.makedirs("chunks", exist_ok=True)
try:
@@ -289,10 +399,22 @@ class Client:
self.p.terminate()
self.close_websocket()
# combine all the audio files
self.write_output_recording(n_audio_file, out_file)
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
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`.
Args:
n_audio_file (int): The number of audio chunk files to combine.
out_file (str): The name of the output WAV file to save the final recording.
"""
input_files = [
f"chunks/{i}.wav"
for i in range(n_audio_file)
@@ -316,10 +438,44 @@ class Client:
class TranscriptionClient:
"""
Client for handling audio transcription tasks via a WebSocket connection.
Acts as a high-level client for audio transcription tasks using a WebSocket connection. It can be used
to send audio data for transcription to a server and receive transcribed text segments.
Args:
host (str): The hostname or IP address of the server.
port (int): The port number to connect to on the server.
is_multilingual (bool, optional): Indicates whether the transcription should support multiple languages (default is False).
lang (str, optional): The primary language for transcription (used if `is_multilingual` is False). Default is None, which defaults to English ('en').
translate (bool, optional): Indicates whether translation tasks are required (default is False).
Attributes:
client (Client): An instance of the underlying Client class responsible for handling the WebSocket connection.
Example:
To create a TranscriptionClient and start transcription on microphone audio:
```python
transcription_client = TranscriptionClient(host="localhost", port=9090, is_multilingual=True)
transcription_client()
```
"""
def __init__(self, host, port, is_multilingual=False, lang=None, translate=False):
self.client = Client(host, port, is_multilingual, lang, translate)
def __call__(self, audio=None):
"""
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
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:
if self.client.waiting:
whisper_live/server.py
+154 -26
View File
@@ -1,16 +1,12 @@
import websockets
import pickle, struct, time, pyaudio
import time
import threading
import os, json
import base64
import wave
import json
import textwrap
import logging
# logging.basicConfig(level = logging.INFO)
from collections import deque
from dataclasses import dataclass
from websockets.sync.server import serve
import torch
@@ -25,35 +21,66 @@ 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.
"""
RATE = 16000
def __init__(self):
# voice activity detection model
self.vad_model = VoiceActivityDetection()
self.vad_threshold = 0.4
self.clients = {}
self.websockets = {}
self.clients_start_time = {}
self.max_clients = 4
self.max_connection_time = 600 # in seconds
self.max_connection_time = 600
def get_wait_time(self):
"""
Calculate and return the estimated wait time for clients.
Returns:
float: The estimated wait time in minutes.
"""
wait_time = None
for k,v in self.clients_start_time.items():
for k, v in self.clients_start_time.items():
current_client_time_remaining = self.max_connection_time - (time.time() - v)
if wait_time is None:
if wait_time is None or current_client_time_remaining < wait_time:
wait_time = current_client_time_remaining
elif current_client_time_remaining < wait_time:
wait_time = current_client_time_remaining
return wait_time/60
return wait_time / 60
def recv_audio(self, websocket):
"""
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
or not. If the audio frame contains speech, it is added to the client's
audio data for ASR.
If the maximum number of clients is reached, the method sends a
"WAIT" status to the client, indicating that they should wait
until a slot is available.
If a client's connection exceeds the maximum allowed time, it will
be disconnected, and the client's resources will be cleaned up.
Args:
websocket (WebSocket): The WebSocket connection for the client.
Raises:
Exception: If there is an error during the audio frame processing.
"""
logging.info("New client connected")
options = websocket.recv()
@@ -63,7 +90,7 @@ class TranscriptionServer:
logging.warning("Client Queue Full. Asking client to wait ...")
wait_time = self.get_wait_time()
response = {
"uid" : options["uid"],
"uid": options["uid"],
"status": "WAIT",
"message": wait_time,
}
@@ -71,7 +98,7 @@ class TranscriptionServer:
websocket.close()
del websocket
return
client = ServeClient(
websocket,
multilingual=options["multilingual"],
@@ -79,9 +106,9 @@ class TranscriptionServer:
task=options["task"],
client_uid=options["uid"]
)
self.clients[websocket] = client
self.clients_start_time[websocket] = time.time()
self.clients_start_time[websocket] = time.time()
while True:
try:
@@ -92,10 +119,11 @@ class TranscriptionServer:
speech_prob = self.vad_model(torch.from_numpy(frame_np.copy()), self.RATE).item()
if speech_prob < self.vad_threshold:
continue
except Exception as e:
logging.error(e)
return
self.clients[websocket].add_frames(frame_np)
elapsed_time = time.time() - self.clients_start_time[websocket]
@@ -109,7 +137,6 @@ class TranscriptionServer:
del websocket
break
except Exception as e:
logging.error(e)
self.clients[websocket].cleanup()
@@ -117,7 +144,6 @@ class TranscriptionServer:
self.clients_start_time.pop(websocket)
logging.info("Connection Closed.")
logging.info(self.clients)
del websocket
break
@@ -134,11 +160,54 @@ class TranscriptionServer:
class ServeClient:
"""
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.
"""
RATE = 16000
SERVER_READY = "SERVER_READY"
DISCONNECT = "DISCONNECT"
def __init__(self, websocket, task="transcribe", device=None, multilingual=False, language=None, client_uid=None):
"""
Initialize a ServeClient instance.
The Whisper model is initialized based on the client's language and device availability.
The transcription thread is started upon initialization. A "SERVER_READY" message is sent
to the client to indicate that the server is ready.
Args:
websocket (WebSocket): The WebSocket connection for the client.
task (str, optional): The task type, e.g., "transcribe." Defaults to "transcribe".
device (str, optional): The device type for Whisper, "cuda" or "cpu". Defaults to None.
multilingual (bool, optional): Whether the client supports multilingual transcription. Defaults to False.
language (str, optional): The language for transcription. Defaults to None.
client_uid (str, optional): A unique identifier for the client. Defaults to None.
"""
self.client_uid = client_uid
self.data = b""
self.frames = b""
@@ -185,14 +254,21 @@ class ServeClient:
def fill_output(self, output):
"""
Format output with current and previous complete segments
into two lines of 50 characters.
Format the current incomplete transcription output by combining it with previous complete segments.
The resulting transcription is wrapped into two lines, each containing a maximum of 50 characters.
It ensures that the combined transcription fits within two lines, with a maximum of 50 characters per line.
Segments are concatenated in the order they exist in the list of previous segments, with the most
recent complete segment first and older segments prepended as needed to maintain the character limit.
If a 3-second pause is detected in the previous segments, any text preceding it is discarded to ensure
the transcription starts with the most recent complete content. The resulting transcription is returned
as a single string.
Args:
output(str): current incomplete segment
output(str): The current incomplete transcription segment.
Returns:
transcription wrapped in two lines
str: A formatted transcription wrapped in two lines.
"""
text = ''
pick_prev = min(len(self.text), self.pick_previous_segments)
@@ -206,6 +282,21 @@ class ServeClient:
return wrapped
def add_frames(self, frame_np):
"""
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.
Args:
frame_np (numpy.ndarray): The audio frame data as a NumPy array.
"""
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):]
@@ -216,7 +307,20 @@ class ServeClient:
def speech_to_text(self):
"""
Process audio stream in an infinite loop.
Process an audio stream in an infinite loop, continuously transcribing the speech.
This method continuously receives audio frames, performs real-time transcription, and sends
transcribed segments to the client via a WebSocket connection.
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
there is no speech for a specified duration to indicate a pause.
Raises:
Exception: If there is an issue with audio processing or WebSocket communication.
"""
# detect language
if self.language is None:
@@ -324,12 +428,21 @@ class ServeClient:
Processes the segments from whisper. Appends all the segments to the list
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
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
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:
transcription for the current chunk
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.
"""
offset = None
self.current_out = ''
@@ -388,6 +501,13 @@ class ServeClient:
return last_segment
def disconnect(self):
"""
Notify the client of disconnection and send a disconnect message.
This method sends a disconnect message to the client via the WebSocket connection to notify them
that the transcription service is disconnecting gracefully.
"""
self.websocket.send(
json.dumps(
{
@@ -398,6 +518,14 @@ class ServeClient:
)
def cleanup(self):
"""
Perform cleanup tasks before exiting the transcription service.
This method performs necessary cleanup tasks, including stopping the transcription thread, marking
the exit flag to indicate the transcription thread should exit gracefully, and destroying resources
associated with the transcription process.
"""
logging.info("Cleaning up.")
self.exit = True
self.transcriber.destroy()