Aaron Boxer
c028c4b584
Add a minimal, non-breaking hook to WhisperLive that allows external projects to post-process transcription segments before they are sent to the client. Changes: - ServeClientBase: add segment_post_processor attribute (default None) - ServeClientBase.send_transcription_to_client: apply post_processor per-segment with error handling (falls back to original segment) - TranscriptionServer: add segment_post_processor parameter to run() and wire it to each client on creation This enables downstream projects to plug in custom processing (e.g. formatting, PII redaction, diarization tagging) without modifying WhisperLive core code.
737 lines
33 KiB
Python
737 lines
33 KiB
Python
import os
|
|
import time
|
|
import threading
|
|
import queue
|
|
import json
|
|
import functools
|
|
import logging
|
|
import shutil
|
|
import tempfile
|
|
from typing import Optional, List
|
|
from fastapi import FastAPI, UploadFile, Form
|
|
from fastapi.middleware.cors import CORSMiddleware
|
|
from starlette.responses import PlainTextResponse, JSONResponse
|
|
import uvicorn
|
|
from faster_whisper import WhisperModel
|
|
import torch
|
|
|
|
from enum import Enum
|
|
|
|
from whisper_live import metrics as wl_metrics
|
|
from typing import List, Optional
|
|
import numpy as np
|
|
from websockets.sync.server import serve
|
|
from websockets.exceptions import ConnectionClosed
|
|
from whisper_live.vad import VoiceActivityDetector
|
|
from whisper_live.backend.base import ServeClientBase
|
|
|
|
logging.basicConfig(level=logging.INFO)
|
|
|
|
class ClientManager:
|
|
def __init__(self, max_clients=4, max_connection_time=600):
|
|
"""
|
|
Initializes the ClientManager with specified limits on client connections and connection durations.
|
|
|
|
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.start_times = {}
|
|
self.max_clients = max_clients
|
|
self.max_connection_time = max_connection_time
|
|
self.lock = threading.Lock()
|
|
|
|
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.
|
|
"""
|
|
with self.lock:
|
|
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.
|
|
"""
|
|
with self.lock:
|
|
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.
|
|
"""
|
|
with self.lock:
|
|
client = self.clients.pop(websocket, None)
|
|
self.start_times.pop(websocket, None)
|
|
if client:
|
|
client.cleanup()
|
|
|
|
def get_wait_time(self):
|
|
"""
|
|
Calculates the estimated wait time for new clients based on the remaining connection times of current clients.
|
|
|
|
Returns:
|
|
The estimated wait time in minutes for new clients to connect. Returns 0 if there are available slots.
|
|
"""
|
|
with self.lock:
|
|
wait_time = None
|
|
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
|
|
|
|
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.
|
|
"""
|
|
with self.lock:
|
|
if len(self.clients) >= self.max_clients:
|
|
wait_time = None
|
|
for start_time in self.start_times.values():
|
|
remaining = self.max_connection_time - (time.time() - start_time)
|
|
if wait_time is None or remaining < wait_time:
|
|
wait_time = remaining
|
|
wait_time_minutes = wait_time / 60 if wait_time is not None else 0
|
|
response = {"uid": options["uid"], "status": "WAIT", "message": wait_time_minutes}
|
|
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.
|
|
"""
|
|
with self.lock:
|
|
elapsed_time = time.time() - self.start_times[websocket]
|
|
client = self.clients.get(websocket)
|
|
if elapsed_time >= self.max_connection_time and client:
|
|
client.disconnect()
|
|
logging.warning(f"Client with uid '{client.client_uid}' disconnected due to overtime.")
|
|
return True
|
|
return False
|
|
|
|
|
|
class BackendType(Enum):
|
|
FASTER_WHISPER = "faster_whisper"
|
|
TENSORRT = "tensorrt"
|
|
OPENVINO = "openvino"
|
|
|
|
@staticmethod
|
|
def valid_types() -> List[str]:
|
|
return [backend_type.value for backend_type in BackendType]
|
|
|
|
@staticmethod
|
|
def is_valid(backend: str) -> bool:
|
|
return backend in BackendType.valid_types()
|
|
|
|
def is_faster_whisper(self) -> bool:
|
|
return self == BackendType.FASTER_WHISPER
|
|
|
|
def is_tensorrt(self) -> bool:
|
|
return self == BackendType.TENSORRT
|
|
|
|
def is_openvino(self) -> bool:
|
|
return self == BackendType.OPENVINO
|
|
|
|
|
|
class TranscriptionServer:
|
|
RATE = 16000
|
|
|
|
def __init__(self):
|
|
self.client_manager = None
|
|
self.no_voice_activity_chunks = 0
|
|
self.use_vad = True
|
|
self.single_model = False
|
|
self.batch_config = None
|
|
self.raw_pcm_input = False
|
|
self.segment_post_processor = None
|
|
|
|
def initialize_client(
|
|
self, websocket, options, faster_whisper_custom_model_path,
|
|
whisper_tensorrt_path, trt_multilingual, trt_py_session=False,
|
|
):
|
|
client: Optional[ServeClientBase] = None
|
|
|
|
# Check if client wants translation
|
|
enable_translation = options.get("enable_translation", False)
|
|
|
|
# Create translation queue if translation is enabled
|
|
translation_queue = None
|
|
translation_client = None
|
|
translation_thread = None
|
|
|
|
if enable_translation:
|
|
target_language = options.get("target_language", "fr")
|
|
translation_queue = queue.Queue(maxsize=ServeClientBase.MAX_TRANSLATION_QUEUE_SIZE)
|
|
from whisper_live.backend.translation_backend import ServeClientTranslation
|
|
translation_client = ServeClientTranslation(
|
|
client_uid=options["uid"],
|
|
websocket=websocket,
|
|
translation_queue=translation_queue,
|
|
target_language=target_language,
|
|
send_last_n_segments=options.get("send_last_n_segments", 10)
|
|
)
|
|
|
|
# Start translation thread
|
|
translation_thread = threading.Thread(
|
|
target=translation_client.speech_to_text,
|
|
daemon=True
|
|
)
|
|
translation_thread.start()
|
|
|
|
logging.info(f"Translation enabled for client {options['uid']} with target language: {target_language}")
|
|
|
|
if self.backend.is_tensorrt():
|
|
try:
|
|
from whisper_live.backend.trt_backend import ServeClientTensorRT
|
|
client = ServeClientTensorRT(
|
|
websocket,
|
|
multilingual=trt_multilingual,
|
|
language=options["language"],
|
|
task=options["task"],
|
|
client_uid=options["uid"],
|
|
model=whisper_tensorrt_path,
|
|
single_model=self.single_model,
|
|
use_py_session=trt_py_session,
|
|
send_last_n_segments=options.get("send_last_n_segments", 10),
|
|
no_speech_thresh=options.get("no_speech_thresh", 0.45),
|
|
clip_audio=options.get("clip_audio", False),
|
|
same_output_threshold=options.get("same_output_threshold", 10),
|
|
)
|
|
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 = BackendType.FASTER_WHISPER
|
|
|
|
if self.backend.is_openvino():
|
|
try:
|
|
from whisper_live.backend.openvino_backend import ServeClientOpenVINO
|
|
client = ServeClientOpenVINO(
|
|
websocket,
|
|
language=options["language"],
|
|
task=options["task"],
|
|
client_uid=options["uid"],
|
|
model=options["model"],
|
|
single_model=self.single_model,
|
|
send_last_n_segments=options.get("send_last_n_segments", 10),
|
|
no_speech_thresh=options.get("no_speech_thresh", 0.45),
|
|
clip_audio=options.get("clip_audio", False),
|
|
same_output_threshold=options.get("same_output_threshold", 10),
|
|
)
|
|
logging.info("Running OpenVINO backend.")
|
|
except Exception as e:
|
|
logging.error(f"OpenVINO not supported: {e}")
|
|
self.backend = BackendType.FASTER_WHISPER
|
|
self.client_uid = options["uid"]
|
|
websocket.send(json.dumps({
|
|
"uid": self.client_uid,
|
|
"status": "WARNING",
|
|
"message": "OpenVINO not supported on Server yet. "
|
|
"Reverting to available backend: 'faster_whisper'"
|
|
}))
|
|
|
|
try:
|
|
if self.backend.is_faster_whisper():
|
|
from whisper_live.backend.faster_whisper_backend import ServeClientFasterWhisper
|
|
# model is of the form namespace/repo_name and not a filesystem path
|
|
if faster_whisper_custom_model_path is not None:
|
|
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"),
|
|
use_vad=self.use_vad,
|
|
single_model=self.single_model,
|
|
send_last_n_segments=options.get("send_last_n_segments", 10),
|
|
no_speech_thresh=options.get("no_speech_thresh", 0.45),
|
|
clip_audio=options.get("clip_audio", False),
|
|
same_output_threshold=options.get("same_output_threshold", 10),
|
|
cache_path=self.cache_path,
|
|
translation_queue=translation_queue,
|
|
hotwords=options.get("hotwords"),
|
|
diarization=self._create_diarizer(options),
|
|
word_timestamps=options.get("word_timestamps", False),
|
|
)
|
|
|
|
logging.info("Running faster_whisper backend.")
|
|
|
|
# Start batch inference worker on first client (after model is loaded)
|
|
if (self.batch_config is not None
|
|
and ServeClientFasterWhisper.BATCH_WORKER is None
|
|
and ServeClientFasterWhisper.SINGLE_MODEL is not None):
|
|
from whisper_live.batch_inference import BatchInferenceWorker
|
|
worker = BatchInferenceWorker(
|
|
transcriber=ServeClientFasterWhisper.SINGLE_MODEL,
|
|
**self.batch_config,
|
|
)
|
|
worker.start()
|
|
ServeClientFasterWhisper.BATCH_WORKER = worker
|
|
except Exception as e:
|
|
logging.error(e)
|
|
return
|
|
|
|
if client is None:
|
|
raise ValueError(f"Backend type {self.backend.value} not recognised or not handled.")
|
|
|
|
# Attach segment post-processor if configured
|
|
if self.segment_post_processor is not None:
|
|
client.segment_post_processor = self.segment_post_processor
|
|
|
|
if translation_client:
|
|
client.translation_client = translation_client
|
|
client.translation_thread = translation_thread
|
|
|
|
self.client_manager.add_client(websocket, client)
|
|
|
|
def _create_diarizer(self, options):
|
|
"""Create a SpeakerDiarizer if the client requested diarization.
|
|
|
|
Returns:
|
|
SpeakerDiarizer or None
|
|
"""
|
|
if not options.get("enable_diarization", False):
|
|
return None
|
|
try:
|
|
from whisper_live.diarization import SpeakerDiarizer
|
|
return SpeakerDiarizer(
|
|
similarity_threshold=options.get("diarization_threshold", 0.55),
|
|
max_speakers=options.get("max_speakers", 10),
|
|
hf_token=options.get("hf_token"),
|
|
)
|
|
except ImportError:
|
|
logging.warning("pyannote.audio not installed; diarization disabled")
|
|
return None
|
|
|
|
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()
|
|
if frame_data == b"END_OF_AUDIO":
|
|
return False
|
|
if self.raw_pcm_input:
|
|
audio_np = np.frombuffer(frame_data, dtype=np.int16)
|
|
return audio_np.astype(np.float32) / 32768.0
|
|
return np.frombuffer(frame_data, dtype=np.float32)
|
|
|
|
def handle_new_connection(self, websocket, faster_whisper_custom_model_path,
|
|
whisper_tensorrt_path, trt_multilingual, trt_py_session=False):
|
|
try:
|
|
logging.info("New client connected")
|
|
options = websocket.recv()
|
|
options = json.loads(options)
|
|
|
|
self.use_vad = options.get('use_vad')
|
|
if self.client_manager.is_server_full(websocket, options):
|
|
wl_metrics.track_connection_rejected(reason="full")
|
|
websocket.close()
|
|
return False # Indicates that the connection should not continue
|
|
|
|
if self.backend.is_tensorrt():
|
|
self.vad_detector = VoiceActivityDetector(frame_rate=self.RATE)
|
|
self.initialize_client(websocket, options, faster_whisper_custom_model_path,
|
|
whisper_tensorrt_path, trt_multilingual, trt_py_session=trt_py_session)
|
|
wl_metrics.track_connection_opened()
|
|
return True
|
|
except json.JSONDecodeError:
|
|
logging.error("Failed to decode JSON from client")
|
|
return False
|
|
except ConnectionClosed:
|
|
logging.info("Connection closed by client")
|
|
return False
|
|
except Exception as e:
|
|
logging.error(f"Error during new connection initialization: {str(e)}")
|
|
return False
|
|
|
|
def process_audio_frames(self, websocket):
|
|
frame_np = self.get_audio_from_websocket(websocket)
|
|
client = self.client_manager.get_client(websocket)
|
|
if frame_np is False:
|
|
if self.backend.is_tensorrt():
|
|
client.set_eos(True)
|
|
return False
|
|
|
|
if self.backend.is_tensorrt():
|
|
voice_active = self.voice_activity(websocket, frame_np)
|
|
if voice_active:
|
|
self.no_voice_activity_chunks = 0
|
|
client.set_eos(False)
|
|
if self.use_vad and not voice_active:
|
|
return True
|
|
|
|
client.add_frames(frame_np)
|
|
return True
|
|
|
|
def recv_audio(self,
|
|
websocket,
|
|
backend: BackendType = BackendType.FASTER_WHISPER,
|
|
faster_whisper_custom_model_path=None,
|
|
whisper_tensorrt_path=None,
|
|
trt_multilingual=False,
|
|
trt_py_session=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
|
|
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.
|
|
backend (str): The backend to run the server with.
|
|
faster_whisper_custom_model_path (str): path to custom faster whisper model.
|
|
whisper_tensorrt_path (str): Required for tensorrt backend.
|
|
trt_multilingual(bool): Only used for tensorrt, True if multilingual model.
|
|
|
|
Raises:
|
|
Exception: If there is an error during the audio frame processing.
|
|
"""
|
|
self.backend = backend
|
|
if not self.handle_new_connection(websocket, faster_whisper_custom_model_path,
|
|
whisper_tensorrt_path, trt_multilingual, trt_py_session=trt_py_session):
|
|
return
|
|
|
|
try:
|
|
while not self.client_manager.is_client_timeout(websocket):
|
|
if not self.process_audio_frames(websocket):
|
|
break
|
|
except ConnectionClosed:
|
|
logging.info("Connection closed by 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()
|
|
wl_metrics.track_connection_closed()
|
|
del websocket
|
|
|
|
def run(self,
|
|
host,
|
|
port=9090,
|
|
backend="tensorrt",
|
|
faster_whisper_custom_model_path=None,
|
|
whisper_tensorrt_path=None,
|
|
trt_multilingual=False,
|
|
trt_py_session=False,
|
|
single_model=False,
|
|
max_clients=4,
|
|
max_connection_time=600,
|
|
cache_path="~/.cache/whisper-live/",
|
|
rest_port=8000,
|
|
enable_rest=False,
|
|
cors_origins: Optional[str] = None,
|
|
batch_enabled=False,
|
|
batch_max_size=8,
|
|
batch_window_ms=50,
|
|
raw_pcm_input=False,
|
|
metrics_port: int = 0,
|
|
segment_post_processor=None):
|
|
"""
|
|
Run the transcription server.
|
|
|
|
Args:
|
|
host (str): The host address to bind the server.
|
|
port (int): The port number to bind the server.
|
|
batch_enabled (bool): Enable cross-client GPU batch inference for
|
|
the faster_whisper backend. When enabled, ``single_model`` is
|
|
forced to True and a ``BatchInferenceWorker`` is started after
|
|
the first client connects. Defaults to False.
|
|
batch_max_size (int): Maximum number of requests per GPU batch.
|
|
Defaults to 8.
|
|
batch_window_ms (int): Maximum time in milliseconds to wait for
|
|
the batch to fill after the first request arrives. Defaults
|
|
to 50.
|
|
segment_post_processor (callable, optional): A callable that receives
|
|
a transcription segment dict and returns a modified segment dict.
|
|
Applied to every segment before sending to the client. Useful for
|
|
plugging in custom post-processing (e.g. formatting, redaction).
|
|
Defaults to None.
|
|
"""
|
|
self.cache_path = cache_path
|
|
self.raw_pcm_input = raw_pcm_input
|
|
|
|
if max_clients < 1:
|
|
raise ValueError(f"max_clients must be >= 1, got {max_clients}")
|
|
if max_connection_time <= 0:
|
|
raise ValueError(f"max_connection_time must be > 0, got {max_connection_time}")
|
|
if batch_enabled and batch_max_size < 1:
|
|
raise ValueError(f"batch_max_size must be >= 1, got {batch_max_size}")
|
|
if batch_enabled and batch_window_ms < 0:
|
|
raise ValueError(f"batch_window_ms must be >= 0, got {batch_window_ms}")
|
|
|
|
self.segment_post_processor = segment_post_processor
|
|
self.client_manager = ClientManager(max_clients, max_connection_time)
|
|
if faster_whisper_custom_model_path is not None and not os.path.exists(faster_whisper_custom_model_path):
|
|
if "/" not in faster_whisper_custom_model_path:
|
|
raise ValueError(f"Custom faster_whisper model '{faster_whisper_custom_model_path}' is not a valid path or HuggingFace model.")
|
|
if whisper_tensorrt_path is not None and not os.path.exists(whisper_tensorrt_path):
|
|
raise ValueError(f"TensorRT model '{whisper_tensorrt_path}' is not a valid path.")
|
|
|
|
# Batch inference config
|
|
if batch_enabled:
|
|
single_model = True # Batch mode requires shared model
|
|
self.batch_config = {
|
|
'max_batch_size': batch_max_size,
|
|
'batch_window_ms': batch_window_ms,
|
|
}
|
|
logging.info(f"Batch inference enabled (max_batch={batch_max_size}, window={batch_window_ms}ms)")
|
|
else:
|
|
self.batch_config = None
|
|
|
|
if single_model:
|
|
if faster_whisper_custom_model_path or whisper_tensorrt_path:
|
|
logging.info("Custom model option was provided. Switching to single model mode.")
|
|
self.single_model = True
|
|
# TODO: load model initially
|
|
else:
|
|
logging.info("Single model mode currently only works with custom models.")
|
|
if not BackendType.is_valid(backend):
|
|
raise ValueError(f"{backend} is not a valid backend type. Choose backend from {BackendType.valid_types()}")
|
|
|
|
# Start Prometheus metrics endpoint if port is specified
|
|
if metrics_port > 0:
|
|
wl_metrics.start_metrics_server(metrics_port)
|
|
|
|
# New OpenAI-compatible REST API (toggleable via enable_rest boolean)
|
|
if enable_rest:
|
|
app = FastAPI(title="WhisperLive OpenAI-Compatible API")
|
|
origins = [o.strip() for o in cors_origins.split(',')] if cors_origins else []
|
|
app.add_middleware(
|
|
CORSMiddleware,
|
|
allow_origins=origins,
|
|
allow_credentials=True,
|
|
allow_methods=["*"], # Allows all methods (GET, POST, etc.)
|
|
allow_headers=["*"], # Allows all headers
|
|
)
|
|
|
|
|
|
@app.post("/v1/audio/transcriptions")
|
|
async def transcribe(
|
|
file: UploadFile,
|
|
model: str = Form(default="whisper-1"),
|
|
language: Optional[str] = Form(default=None),
|
|
prompt: Optional[str] = Form(default=None),
|
|
response_format: str = Form(default="json"),
|
|
temperature: float = Form(default=0.0),
|
|
timestamp_granularities: Optional[List[str]] = Form(default=None),
|
|
# Stubs for unsupported OpenAI params
|
|
chunking_strategy: Optional[str] = Form(default=None),
|
|
include: Optional[List[str]] = Form(default=None),
|
|
known_speaker_names: Optional[List[str]] = Form(default=None),
|
|
known_speaker_references: Optional[List[str]] = Form(default=None),
|
|
stream: bool = Form(default=False),
|
|
hotwords: Optional[str] = Form(default=None),
|
|
):
|
|
if stream:
|
|
wl_metrics.track_rest_request(endpoint="transcriptions", status=400)
|
|
return JSONResponse({"error": "Streaming not supported in this backend."}, status_code=400)
|
|
if chunking_strategy or known_speaker_names or known_speaker_references:
|
|
logging.warning("Diarization/chunking params ignored; not supported.")
|
|
|
|
supported_formats = ["json", "text", "srt", "verbose_json", "vtt"]
|
|
if response_format not in supported_formats:
|
|
wl_metrics.track_rest_request(endpoint="transcriptions", status=400)
|
|
return JSONResponse({"error": f"Unsupported response_format. Supported: {supported_formats}"}, status_code=400)
|
|
|
|
if model != "whisper-1":
|
|
logging.warning(f"Model '{model}' requested; using 'small' as fallback.")
|
|
model_name = faster_whisper_custom_model_path or "small"
|
|
|
|
try:
|
|
suffix = os.path.splitext(file.filename)[1] or ".wav"
|
|
with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
|
|
shutil.copyfileobj(file.file, tmp)
|
|
tmp_path = tmp.name
|
|
|
|
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
compute_type = "float16" if device == "cuda" else "int8"
|
|
|
|
transcriber = WhisperModel(model_name, device=device, compute_type=compute_type)
|
|
segments, info = transcriber.transcribe(
|
|
tmp_path,
|
|
language=language,
|
|
initial_prompt=prompt,
|
|
temperature=temperature,
|
|
vad_filter=False,
|
|
word_timestamps=(timestamp_granularities and "word" in timestamp_granularities),
|
|
hotwords=hotwords,
|
|
)
|
|
|
|
text = " ".join([s.text.strip() for s in segments])
|
|
os.unlink(tmp_path)
|
|
|
|
if response_format == "text":
|
|
wl_metrics.track_rest_request(endpoint="transcriptions", status=200)
|
|
return PlainTextResponse(text)
|
|
elif response_format == "json":
|
|
wl_metrics.track_rest_request(endpoint="transcriptions", status=200)
|
|
return {"text": text}
|
|
elif response_format == "verbose_json":
|
|
verbose = {
|
|
"task": "transcribe",
|
|
"language": info.language,
|
|
"duration": info.duration,
|
|
"text": text,
|
|
"segments": []
|
|
}
|
|
for seg in segments:
|
|
seg_dict = {
|
|
"id": seg.id,
|
|
"seek": seg.seek,
|
|
"start": seg.start,
|
|
"end": seg.end,
|
|
"text": seg.text.strip(),
|
|
"tokens": seg.tokens,
|
|
"temperature": seg.temperature,
|
|
"avg_logprob": seg.avg_logprob,
|
|
"compression_ratio": seg.compression_ratio,
|
|
"no_speech_prob": seg.no_speech_prob
|
|
}
|
|
if timestamp_granularities and "word" in timestamp_granularities:
|
|
seg_dict["words"] = [{"word": w.word, "start": w.start, "end": w.end, "probability": w.probability} for w in seg.words]
|
|
verbose["segments"].append(seg_dict)
|
|
wl_metrics.track_rest_request(endpoint="transcriptions", status=200)
|
|
return verbose
|
|
elif response_format in ["srt", "vtt"]:
|
|
output = []
|
|
for i, seg in enumerate(segments, 1):
|
|
start = f"{int(seg.start // 3600):02}:{int((seg.start % 3600) // 60):02}:{seg.start % 60:06.3f}"
|
|
end = f"{int(seg.end // 3600):02}:{int((seg.end % 3600) // 60):02}:{seg.end % 60:06.3f}"
|
|
if response_format == "srt":
|
|
output.append(f"{i}\n{start.replace('.', ',')} --> {end.replace('.', ',')}\n{seg.text.strip()}\n")
|
|
else: # vtt
|
|
output.append(f"{start} --> {end}\n{seg.text.strip()}\n")
|
|
wl_metrics.track_rest_request(endpoint="transcriptions", status=200)
|
|
return PlainTextResponse("\n".join(output))
|
|
except Exception as e:
|
|
wl_metrics.track_rest_request(endpoint="transcriptions", status=500)
|
|
wl_metrics.track_error("rest_transcription")
|
|
return JSONResponse({"error": str(e)}, status_code=500)
|
|
|
|
threading.Thread(
|
|
target=uvicorn.run,
|
|
args=(app,),
|
|
kwargs={"host": "0.0.0.0", "port": rest_port, "log_level": "info"},
|
|
daemon=True
|
|
).start()
|
|
logging.info(f"✅ OpenAI-Compatible API started on http://0.0.0.0:{rest_port}")
|
|
|
|
# Original WebSocket server (always supported)
|
|
with serve(
|
|
functools.partial(
|
|
self.recv_audio,
|
|
backend=BackendType(backend),
|
|
faster_whisper_custom_model_path=faster_whisper_custom_model_path,
|
|
whisper_tensorrt_path=whisper_tensorrt_path,
|
|
trt_multilingual=trt_multilingual,
|
|
trt_py_session=trt_py_session,
|
|
),
|
|
host,
|
|
port
|
|
) 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.
|
|
"""
|
|
client = self.client_manager.get_client(websocket)
|
|
if client:
|
|
if hasattr(client, 'translation_client') and client.translation_client:
|
|
client.translation_client.cleanup()
|
|
|
|
# Wait for translation thread to finish
|
|
if hasattr(client, 'translation_thread') and client.translation_thread:
|
|
client.translation_thread.join(timeout=2.0)
|
|
self.client_manager.remove_client(websocket) |