update faster-whisper

Dockerfile

@@ -1,4 +1,4 @@
-FROM docker.io/nvidia/cuda:12.0.0-cudnn8-runtime-ubuntu22.04
+FROM nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04
 
 RUN apt-get update && \
     apt-get install -y ffmpeg python3 python3-pip git && \

README.md

@@ -5,7 +5,7 @@ Whisper-FastAPI is a very simple Python FastAPI interface for konele and OpenAI
 ## Features
 
 - **Translation and Transcription**: The application provides an API for konele service, where translations and transcriptions can be obtained by connecting over websockets or POST requests.
-- **Language Support**: If the target language is English, then the application will translate any source language to English.
+- **Language Support**: If no language is specified, the language will be automatically recognized from the first 30 seconds.
 - **Websocket and POST Method Support**: The project supports a websocket (`/konele/ws`) and a POST method to `/konele/post`.
 - **Audio Transcriptions**: The `/v1/audio/transcriptions` endpoint allows users to upload an audio file and receive transcription in response, with an optional `response_type` parameter. The `response_type` can be 'json', 'text', 'tsv', 'srt', and 'vtt'.
 - **Simplified Chinese**: The traditional Chinese will be automatically convert to simplified Chinese for konele using `opencc` library.
@@ -16,10 +16,10 @@ Whisper-FastAPI is a very simple Python FastAPI interface for konele and OpenAI
 
 For konele voice typing, you can use either the websocket endpoint or the POST method endpoint.
 
-- **Websocket**: Connect to the websocket at `/konele/ws` and send audio data. The server will respond with the transcription or translation.
-- **POST Method**: Send a POST request to `/konele/post` with the audio data in the body. The server will respond with the transcription or translation.
+- **Websocket**: Connect to the websocket at `/konele/ws` (or `/v1/konele/ws`) and send audio data. The server will respond with the transcription or translation.
+- **POST Method**: Send a POST request to `/konele/post` (or `/v1/konele/post`) with the audio data in the body. The server will respond with the transcription or translation.
 
-You can also use the demo I have created to quickly test the effect at <https://yongyuancv.cn/konele/ws> and <https://yongyuancv.cn/konele/post>
+You can also use the demo I have created to quickly test the effect at <https://yongyuancv.cn/v1/konele/ws> and <https://yongyuancv.cn/v1/konele/post>
 
 ### OpenAI Whisper Service
 

requirements.txt

@@ -4,5 +4,5 @@ uvicorn[standard]
 whisper_ctranslate2
 opencc
 prometheus-fastapi-instrumentator
-git+https://github.com/heimoshuiyu/faster-whisper@prompt
+git+https://github.com/heimoshuiyu/faster-whisper@a759f5f48f5ef5b79461a6461966eafe9df088a9
 pydub

requirements_version.txt

@@ -1,49 +1,49 @@
 annotated-types==0.7.0
-anyio==4.4.0
-av==12.3.0
+anyio==4.6.2.post1
+av==13.1.0
 certifi==2024.8.30
 cffi==1.17.1
-charset-normalizer==3.3.2
+charset-normalizer==3.4.0
 click==8.1.7
 coloredlogs==15.0.1
-ctranslate2==4.4.0
+ctranslate2==4.5.0
 exceptiongroup==1.2.2
-fastapi==0.114.1
-faster-whisper @ git+https://github.com/heimoshuiyu/faster-whisper@28a4d11a736d8cdeb4655ee5d7e4b4e7ae5ec8e0
-filelock==3.16.0
+fastapi==0.115.5
+faster-whisper @ git+https://github.com/heimoshuiyu/faster-whisper@a759f5f48f5ef5b79461a6461966eafe9df088a9
+filelock==3.16.1
 flatbuffers==24.3.25
-fsspec==2024.9.0
+fsspec==2024.10.0
 h11==0.14.0
-httptools==0.6.1
-huggingface-hub==0.24.6
+httptools==0.6.4
+huggingface-hub==0.26.2
 humanfriendly==10.0
-idna==3.8
+idna==3.10
 mpmath==1.3.0
-numpy==2.1.1
-onnxruntime==1.19.2
+numpy==2.1.3
+onnxruntime==1.20.1
 OpenCC==1.1.9
-packaging==24.1
+packaging==24.2
 prometheus-fastapi-instrumentator==7.0.0
-prometheus_client==0.20.0
-protobuf==5.28.0
+prometheus_client==0.21.0
+protobuf==5.28.3
 pycparser==2.22
-pydantic==2.9.1
-pydantic_core==2.23.3
+pydantic==2.10.1
+pydantic_core==2.27.1
 pydub==0.25.1
 python-dotenv==1.0.1
-python-multipart==0.0.9
+python-multipart==0.0.17
 PyYAML==6.0.2
 requests==2.32.3
 sniffio==1.3.1
-sounddevice==0.5.0
-starlette==0.38.5
-sympy==1.13.2
-tokenizers==0.20.0
-tqdm==4.66.5
+sounddevice==0.5.1
+starlette==0.41.3
+sympy==1.13.3
+tokenizers==0.20.3
+tqdm==4.67.0
 typing_extensions==4.12.2
-urllib3==2.2.2
-uvicorn==0.30.6
-uvloop==0.20.0
+urllib3==2.2.3
+uvicorn==0.32.1
+uvloop==0.21.0
 watchfiles==0.24.0
-websockets==13.0.1
-whisper-ctranslate2==0.4.5
+websockets==14.1
+whisper-ctranslate2==0.4.8

whisper_fastapi.py

@@ -1,4 +1,6 @@
-import tqdm
+import sys
+import dataclasses
+import faster_whisper
 import json
 from fastapi.responses import StreamingResponse
 import wave
@@ -7,7 +9,7 @@ import io
 import hashlib
 import argparse
 import uvicorn
-from typing import Annotated, Any, BinaryIO, Literal, Generator, Tuple, Iterable
+from typing import Annotated, Any, BinaryIO, Literal, Generator, Tuple, Iterable, Union
 from fastapi import (
     File,
     HTTPException,
@@ -19,12 +21,16 @@ from fastapi import (
     WebSocket,
 )
 from fastapi.middleware.cors import CORSMiddleware
-from src.whisper_ctranslate2.whisper_ctranslate2 import Transcribe
 from src.whisper_ctranslate2.writers import format_timestamp
 from faster_whisper.transcribe import Segment, TranscriptionInfo
 import opencc
 from prometheus_fastapi_instrumentator import Instrumentator
 
+# redirect print to stderr
+_print = print
+def print(*args, **kwargs):
+    _print(*args, file=sys.stderr, **kwargs)
+
 parser = argparse.ArgumentParser()
 parser.add_argument("--host", default="0.0.0.0", type=str)
 parser.add_argument("--port", default=5000, type=int)
@@ -40,16 +46,13 @@ Instrumentator().instrument(app).expose(app, endpoint="/konele/metrics")
 ccc = opencc.OpenCC("t2s.json")
 
 print(f"Loading model to device {args.device}...")
-transcriber = Transcribe(
-    model_path=args.model,
+model = faster_whisper.WhisperModel(
+    model_size_or_path=args.model,
     device=args.device,
-    device_index=0,
-    compute_type="default",
-    threads=args.threads,
-    cache_directory=args.cache_dir,
+    cpu_threads=args.threads,
     local_files_only=args.local_files_only,
 )
-print(f"Model loaded to device {transcriber.model.model.device}")
+print(f"Model loaded to device {model.model.device}")
 
 
 # allow all cors
@@ -62,56 +65,63 @@ app.add_middleware(
 )
 
 
-def stream_writer(generator: Generator[dict[str, Any], Any, None]):
+def stream_writer(generator: Generator[Segment, Any, None]):
     for segment in generator:
         yield "data: " + json.dumps(segment, ensure_ascii=False) + "\n\n"
     yield "data: [DONE]\n\n"
 
 
-def text_writer(generator: Generator[dict[str, Any], Any, None]):
+def text_writer(generator: Generator[Segment, Any, None]):
     for segment in generator:
-        yield segment["text"].strip() + "\n"
+        yield segment.text.strip() + "\n"
 
 
-def tsv_writer(generator: Generator[dict[str, Any], Any, None]):
+def tsv_writer(generator: Generator[Segment, Any, None]):
     yield "start\tend\ttext\n"
     for i, segment in enumerate(generator):
-        start_time = str(round(1000 * segment["start"]))
-        end_time = str(round(1000 * segment["end"]))
-        text = segment["text"].strip()
+        start_time = str(round(1000 * segment.start))
+        end_time = str(round(1000 * segment.end))
+        text = segment.text.strip()
         yield f"{start_time}\t{end_time}\t{text}\n"
 
 
-def srt_writer(generator: Generator[dict[str, Any], Any, None]):
+def srt_writer(generator: Generator[Segment, Any, None]):
     for i, segment in enumerate(generator):
         start_time = format_timestamp(
-            segment["start"], decimal_marker=",", always_include_hours=True
+            segment.start, decimal_marker=",", always_include_hours=True
         )
         end_time = format_timestamp(
-            segment["end"], decimal_marker=",", always_include_hours=True
+            segment.end, decimal_marker=",", always_include_hours=True
         )
-        text = segment["text"].strip()
+        text = segment.text.strip()
         yield f"{i}\n{start_time} --> {end_time}\n{text}\n\n"
 
 
-def vtt_writer(generator: Generator[dict[str, Any], Any, None]):
+def vtt_writer(generator: Generator[Segment, Any, None]):
     yield "WEBVTT\n\n"
-    for i, segment in enumerate(generator):
-        start_time = format_timestamp(segment["start"])
-        end_time = format_timestamp(segment["end"])
-        text = segment["text"].strip()
+    for _, segment in enumerate(generator):
+        start_time = format_timestamp(segment.start)
+        end_time = format_timestamp(segment.end)
+        text = segment.text.strip()
         yield f"{start_time} --> {end_time}\n{text}\n\n"
 
 
+@dataclasses.dataclass
+class JsonResult(TranscriptionInfo):
+    segments: list[Segment]
+    text: str
+
+
 def build_json_result(
     generator: Iterable[Segment],
-    info:  dict,
-) -> dict[str, Any]:
+    info: TranscriptionInfo,
+) -> JsonResult:
     segments = [i for i in generator]
-    return info | {
-        "text": "\n".join(i["text"] for i in segments),
-        "segments": segments,
-    }
+    return JsonResult(
+        text="\n".join(i.text for i in segments),
+        segments=segments,
+        **dataclasses.asdict(info),
+    )
 
 
 def stream_builder(
@@ -121,12 +131,13 @@ def stream_builder(
     language: str | None,
     initial_prompt: str = "",
     repetition_penalty: float = 1.0,
-) -> Tuple[Iterable[dict], dict]:
-    segments, info = transcriber.model.transcribe(
+) -> Tuple[Generator[Segment, None, None], TranscriptionInfo]:
+    segments, info = model.transcribe(
         audio=audio,
         language=language,
         task=task,
-        initial_prompt=initial_prompt,
+        vad_filter=vad_filter,
+        initial_prompt=initial_prompt if initial_prompt else None,
         word_timestamps=True,
         repetition_penalty=repetition_penalty,
     )
@@ -134,27 +145,14 @@ def stream_builder(
         "Detected language '%s' with probability %f"
         % (info.language, info.language_probability)
     )
-    def wrap():
-        last_pos = 0
-        with tqdm.tqdm(total=info.duration, unit="seconds", disable=True) as pbar:
-            for segment in segments:
-                start, end, text = segment.start, segment.end, segment.text
-                pbar.update(end - last_pos)
-                last_pos = end
-                data = segment._asdict()
-                if data.get('words') is not None:
-                    data["words"] = [i._asdict() for i in data["words"]]
-                if info.language == "zh":
-                    data["text"] = ccc.convert(data["text"])
-                yield data
 
-    info_dict = info._asdict()
-    if info_dict['transcription_options'] is not None:
-        info_dict['transcription_options'] = info_dict['transcription_options']._asdict()
-    if info_dict['vad_options'] is not None:
-        info_dict['vad_options'] = info_dict['vad_options']._asdict()
-    
-    return wrap(), info_dict
+    def wrap():
+        for segment in segments:
+            if info.language == "zh":
+                segment.text = ccc.convert(segment.text)
+            yield segment
+
+    return wrap(), info
 
 
 @app.websocket("/k6nele/status")
@@ -222,13 +220,11 @@ async def konele_ws(
     )
     result = build_json_result(generator, info)
 
-    text = result.get("text", "")
-
     await websocket.send_json(
         {
             "status": 0,
             "segment": 0,
-            "result": {"hypotheses": [{"transcript": text}], "final": True},
+            "result": {"hypotheses": [{"transcript": result.text}], "final": True},
             "id": md5,
         }
     )
@@ -285,21 +281,21 @@ async def translateapi(
     )
     result = build_json_result(generator, info)
 
-    text = result.get("text", "")
-
     return {
         "status": 0,
-        "hypotheses": [{"utterance": text}],
+        "hypotheses": [{"utterance": result.text}],
         "id": md5,
     }
 
 
-@app.post("/v1/audio/transcriptions")
+@app.post("/v1/audio/transcriptions", response_model=Union[JsonResult, str])
+@app.post("/v1/audio/translations", response_model=Union[JsonResult, str])
 async def transcription(
+    request: Request,
     file: UploadFile = File(...),
     prompt: str = Form(""),
     response_format: str = Form("json"),
-    task: str = Form("transcribe"),
+    task: str = Form(""),
     language: str = Form("und"),
     vad_filter: bool = Form(False),
     repetition_penalty: float = Form(1.0),
@@ -309,11 +305,20 @@ async def transcription(
     User upload audio file in multipart/form-data format and receive transcription in response
     """
 
+    if not task:
+        if request.url.path == "/v1/audio/transcriptions":
+            task = "transcribe"
+        elif request.url.path == "/v1/audio/translations":
+            task = "translate"
+        else:
+            raise HTTPException(400, "task parameter is required")
+
     # timestamp as filename, keep original extension
     generator, info = stream_builder(
         audio=io.BytesIO(file.file.read()),
         task=task,
         vad_filter=vad_filter,
+        initial_prompt=prompt,
         language=None if language == "und" else language,
         repetition_penalty=repetition_penalty,
     )