Integrate OpenVINO backend
Signed-off-by: makaveli <vineet.suryan@collabora.com>
This commit is contained in:
makaveli
@@ -1,6 +1,7 @@
|
||||
import json
|
||||
import logging
|
||||
import threading
|
||||
import time
|
||||
import numpy as np
|
||||
|
||||
|
||||
@@ -26,6 +27,7 @@ class ServeClientBase(object):
|
||||
self.add_pause_thresh = 3 # add a blank to segment list as a pause(no speech) for 3 seconds
|
||||
self.transcript = []
|
||||
self.send_last_n_segments = 10
|
||||
self.no_speech_thresh = 0.45
|
||||
|
||||
# text formatting
|
||||
self.pick_previous_segments = 2
|
||||
@@ -34,13 +36,76 @@ class ServeClientBase(object):
|
||||
self.lock = threading.Lock()
|
||||
|
||||
def speech_to_text(self):
|
||||
raise NotImplementedError
|
||||
"""
|
||||
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.
|
||||
|
||||
"""
|
||||
while True:
|
||||
if self.exit:
|
||||
logging.info("Exiting speech to text thread")
|
||||
break
|
||||
|
||||
if self.frames_np is None:
|
||||
continue
|
||||
|
||||
self.clip_audio_if_no_valid_segment()
|
||||
|
||||
input_bytes, duration = self.get_audio_chunk_for_processing()
|
||||
if duration < 1.0:
|
||||
time.sleep(0.1) # wait for audio chunks to arrive
|
||||
continue
|
||||
try:
|
||||
input_sample = input_bytes.copy()
|
||||
result = self.transcribe_audio(input_sample)
|
||||
|
||||
if result is None or self.language is None:
|
||||
self.timestamp_offset += duration
|
||||
time.sleep(0.25) # wait for voice activity, result is None when no voice activity
|
||||
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 transcribe_audio(self):
|
||||
raise NotImplementedError
|
||||
|
||||
def handle_transcription_output(self):
|
||||
raise NotImplementedError
|
||||
|
||||
def format_segment(self, start, end, text, completed=False):
|
||||
"""
|
||||
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),
|
||||
'text': text,
|
||||
'completed': completed
|
||||
}
|
||||
|
||||
def add_frames(self, frame_np):
|
||||
"""
|
||||
@@ -161,6 +226,33 @@ class ServeClientBase(object):
|
||||
except Exception as e:
|
||||
logging.error(f"[ERROR]: Sending data to client: {e}")
|
||||
|
||||
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 disconnect(self):
|
||||
"""
|
||||
Notify the client of disconnection and send a disconnect message.
|
||||
@@ -185,4 +277,94 @@ class ServeClientBase(object):
|
||||
"""
|
||||
logging.info("Cleaning up.")
|
||||
self.exit = True
|
||||
|
||||
def get_segment_no_speech_prob(self, segment):
|
||||
return getattr(segment, "no_speech_prob", 0)
|
||||
|
||||
def get_segment_start(self, segment):
|
||||
return getattr(segment, "start", getattr(segment, "start_ts", 0))
|
||||
|
||||
def get_segment_end(self, segment):
|
||||
return getattr(segment, "end", getattr(segment, "end_ts", 0))
|
||||
|
||||
def update_segments(self, segments, duration):
|
||||
"""
|
||||
Processes the segments from Whisper and updates the transcript.
|
||||
Uses helper methods to account for differences between backends.
|
||||
|
||||
Args:
|
||||
segments (list): List of segments returned by the transcriber.
|
||||
duration (float): Duration of the current audio chunk.
|
||||
|
||||
Returns:
|
||||
dict or None: The last processed segment (if any).
|
||||
"""
|
||||
offset = None
|
||||
self.current_out = ''
|
||||
last_segment = None
|
||||
|
||||
# Process complete segments only if there are more than one
|
||||
# and if the last segment's no_speech_prob is below the threshold.
|
||||
if len(segments) > 1 and self.get_segment_no_speech_prob(segments[-1]) <= self.no_speech_thresh:
|
||||
for s in segments[:-1]:
|
||||
text_ = s.text
|
||||
self.text.append(text_)
|
||||
with self.lock:
|
||||
start = self.timestamp_offset + self.get_segment_start(s)
|
||||
end = self.timestamp_offset + min(duration, self.get_segment_end(s))
|
||||
if start >= end:
|
||||
continue
|
||||
if self.get_segment_no_speech_prob(s) > self.no_speech_thresh:
|
||||
continue
|
||||
self.transcript.append(self.format_segment(start, end, text_, completed=True))
|
||||
offset = min(duration, self.get_segment_end(s))
|
||||
|
||||
# Process the last segment if its no_speech_prob is acceptable.
|
||||
if self.get_segment_no_speech_prob(segments[-1]) <= self.no_speech_thresh:
|
||||
self.current_out += segments[-1].text
|
||||
with self.lock:
|
||||
last_segment = self.format_segment(
|
||||
self.timestamp_offset + self.get_segment_start(segments[-1]),
|
||||
self.timestamp_offset + min(duration, self.get_segment_end(segments[-1])),
|
||||
self.current_out,
|
||||
completed=False
|
||||
)
|
||||
|
||||
# Handle repeated output logic.
|
||||
if self.current_out.strip() == self.prev_out.strip() and self.current_out != '':
|
||||
self.same_output_count += 1
|
||||
|
||||
# if we remove the audio because of same output on the nth reptition we might remove the
|
||||
# audio thats not yet transcribed so, capturing the time when it was repeated for the first time
|
||||
if self.end_time_for_same_output is None:
|
||||
self.end_time_for_same_output = self.get_segment_end(segments[-1])
|
||||
time.sleep(0.1) # wait briefly for any new voice activity
|
||||
else:
|
||||
self.same_output_count = 0
|
||||
self.end_time_for_same_output = None
|
||||
|
||||
# If the same incomplete segment is repeated too many times,
|
||||
# append it to the transcript and update the offset.
|
||||
if self.same_output_count > self.same_output_threshold:
|
||||
if not self.text or self.text[-1].strip().lower() != self.current_out.strip().lower():
|
||||
self.text.append(self.current_out)
|
||||
with self.lock:
|
||||
self.transcript.append(self.format_segment(
|
||||
self.timestamp_offset,
|
||||
self.timestamp_offset + min(duration, self.end_time_for_same_output),
|
||||
self.current_out,
|
||||
completed=True
|
||||
))
|
||||
self.current_out = ''
|
||||
offset = min(duration, self.end_time_for_same_output)
|
||||
self.same_output_count = 0
|
||||
last_segment = None
|
||||
self.end_time_for_same_output = None
|
||||
else:
|
||||
self.prev_out = self.current_out
|
||||
|
||||
if offset is not None:
|
||||
with self.lock:
|
||||
self.timestamp_offset += offset
|
||||
|
||||
return last_segment
|
||||
|
||||
@@ -44,7 +44,7 @@ class ServeClientFasterWhisper(ServeClientBase):
|
||||
self.task = task
|
||||
self.initial_prompt = initial_prompt
|
||||
self.vad_parameters = vad_parameters or {"onset": 0.5}
|
||||
self.no_speech_thresh = 0.45
|
||||
|
||||
self.same_output_threshold = 10
|
||||
self.end_time_for_same_output = None
|
||||
|
||||
@@ -221,161 +221,3 @@ class ServeClientFasterWhisper(ServeClientBase):
|
||||
|
||||
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.
|
||||
|
||||
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.
|
||||
|
||||
"""
|
||||
while True:
|
||||
if self.exit:
|
||||
logging.info("Exiting speech to text thread")
|
||||
break
|
||||
|
||||
if self.frames_np is None:
|
||||
continue
|
||||
|
||||
self.clip_audio_if_no_valid_segment()
|
||||
|
||||
input_bytes, duration = self.get_audio_chunk_for_processing()
|
||||
if duration < 1.0:
|
||||
time.sleep(0.1) # wait for audio chunks to arrive
|
||||
continue
|
||||
try:
|
||||
input_sample = input_bytes.copy()
|
||||
result = self.transcribe_audio(input_sample)
|
||||
|
||||
if result is None or self.language is None:
|
||||
self.timestamp_offset += duration
|
||||
time.sleep(0.25) # wait for voice activity, result is None when no voice activity
|
||||
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, completed=False):
|
||||
"""
|
||||
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),
|
||||
'text': text,
|
||||
'completed': completed
|
||||
}
|
||||
|
||||
def update_segments(self, segments, duration):
|
||||
"""
|
||||
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:
|
||||
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 = ''
|
||||
last_segment = None
|
||||
|
||||
# process complete segments
|
||||
if len(segments) > 1 and segments[-1].no_speech_prob <= self.no_speech_thresh:
|
||||
for i, s in enumerate(segments[:-1]):
|
||||
text_ = s.text
|
||||
self.text.append(text_)
|
||||
with self.lock:
|
||||
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
|
||||
|
||||
self.transcript.append(self.format_segment(start, end, text_, completed=True))
|
||||
offset = min(duration, s.end)
|
||||
|
||||
# only process the last segment if it satisfies the no_speech_thresh
|
||||
if segments[-1].no_speech_prob <= self.no_speech_thresh:
|
||||
self.current_out += segments[-1].text
|
||||
with self.lock:
|
||||
last_segment = self.format_segment(
|
||||
self.timestamp_offset + segments[-1].start,
|
||||
self.timestamp_offset + min(duration, segments[-1].end),
|
||||
self.current_out,
|
||||
completed=False
|
||||
)
|
||||
|
||||
if self.current_out.strip() == self.prev_out.strip() and self.current_out != '':
|
||||
self.same_output_count += 1
|
||||
|
||||
# if we remove the audio because of same output on the nth reptition we might remove the
|
||||
# audio thats not yet transcribed so, capturing the time when it was repeated for the first time
|
||||
if self.end_time_for_same_output is None:
|
||||
self.end_time_for_same_output = segments[-1].end
|
||||
time.sleep(0.1) # wait for some voice activity just in case there is an unitended pause from the speaker for better punctuations.
|
||||
else:
|
||||
self.same_output_count = 0
|
||||
self.end_time_for_same_output = None
|
||||
|
||||
# if same incomplete segment is seen multiple times then update the offset
|
||||
# and append the segment to the list
|
||||
if self.same_output_count > self.same_output_threshold:
|
||||
if not len(self.text) or self.text[-1].strip().lower() != self.current_out.strip().lower():
|
||||
self.text.append(self.current_out)
|
||||
with self.lock:
|
||||
self.transcript.append(self.format_segment(
|
||||
self.timestamp_offset,
|
||||
self.timestamp_offset + min(duration, self.end_time_for_same_output),
|
||||
self.current_out,
|
||||
completed=True
|
||||
))
|
||||
self.current_out = ''
|
||||
offset = min(duration, self.end_time_for_same_output)
|
||||
self.same_output_count = 0
|
||||
last_segment = None
|
||||
self.end_time_for_same_output = None
|
||||
else:
|
||||
self.prev_out = self.current_out
|
||||
|
||||
# update offset
|
||||
if offset is not None:
|
||||
with self.lock:
|
||||
self.timestamp_offset += offset
|
||||
|
||||
return last_segment
|
||||
|
||||
|
||||
@@ -0,0 +1,125 @@
|
||||
import json
|
||||
import logging
|
||||
import threading
|
||||
import time
|
||||
|
||||
from openvino import Core
|
||||
from whisper_live.backend.base import ServeClientBase
|
||||
from whisper_live.transcriber.transcriber_openvino import WhisperOpenVINO
|
||||
|
||||
|
||||
class ServeClientOpenVINO(ServeClientBase):
|
||||
SINGLE_MODEL = None
|
||||
SINGLE_MODEL_LOCK = threading.Lock()
|
||||
|
||||
def __init__(self, websocket, task="transcribe", device=None, language=None, client_uid=None, model="small.en",
|
||||
initial_prompt=None, vad_parameters=None, use_vad=True, single_model=False):
|
||||
"""
|
||||
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.
|
||||
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): Huggingface model_id for a valid OpenVINO model.
|
||||
initial_prompt (str, optional): Prompt for whisper inference. Defaults to None.
|
||||
single_model (bool, optional): Whether to instantiate a new model for each client connection. Defaults to False.
|
||||
"""
|
||||
super().__init__(client_uid, websocket)
|
||||
self.language = "en" if language is None else language
|
||||
if not self.language.startswith("<|"):
|
||||
self.language = f"<|{self.language}|>"
|
||||
|
||||
self.task = "transcribe" if task is None else task
|
||||
self.same_output_threshold = 10
|
||||
self.end_time_for_same_output = None
|
||||
|
||||
core = Core()
|
||||
available_devices = core.available_devices
|
||||
if 'GPU' in available_devices:
|
||||
selected_device = 'GPU'
|
||||
else:
|
||||
gpu_devices = [d for d in available_devices if d.startswith('GPU')]
|
||||
selected_device = gpu_devices[0] if gpu_devices else 'CPU'
|
||||
self.device = selected_device
|
||||
|
||||
|
||||
if single_model:
|
||||
if ServeClientOpenVINO.SINGLE_MODEL is None:
|
||||
self.create_model(model)
|
||||
ServeClientOpenVINO.SINGLE_MODEL = self.transcriber
|
||||
else:
|
||||
self.transcriber = ServeClientOpenVINO.SINGLE_MODEL
|
||||
else:
|
||||
self.create_model(model)
|
||||
|
||||
# threading
|
||||
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": "openvino"
|
||||
}))
|
||||
logging.info(f"Using OpenVINO device: {self.device}")
|
||||
logging.info(f"Running OpenVINO backend with language: {self.language} and task: {self.task}")
|
||||
|
||||
def create_model(self, model_id):
|
||||
"""
|
||||
Instantiates a new model, sets it as the transcriber.
|
||||
"""
|
||||
self.transcriber = WhisperOpenVINO(
|
||||
model_id,
|
||||
device=self.device,
|
||||
language=self.language,
|
||||
task=self.task
|
||||
)
|
||||
|
||||
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.
|
||||
"""
|
||||
if ServeClientOpenVINO.SINGLE_MODEL:
|
||||
ServeClientOpenVINO.SINGLE_MODEL_LOCK.acquire()
|
||||
result = self.transcriber.transcribe(input_sample)
|
||||
if ServeClientOpenVINO.SINGLE_MODEL:
|
||||
ServeClientOpenVINO.SINGLE_MODEL_LOCK.release()
|
||||
return result
|
||||
|
||||
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)
|
||||
Reference in New Issue
Block a user