revert faster-whisper to v1.0.3

.dockerignore

@@ -0,0 +1 @@
+/venv

Dockerfile

@@ -0,0 +1,19 @@
+FROM docker.io/nvidia/cuda:12.0.0-cudnn8-runtime-ubuntu22.04
+
+RUN apt-get update && \
+    apt-get install -y ffmpeg python3 python3-pip git && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+COPY requirements.txt .
+
+RUN pip3 install --no-cache-dir -r requirements.txt
+
+COPY . .
+
+EXPOSE 5000
+
+# 启动 whisper_fastapi.py
+ENTRYPOINT ["python3", "whisper_fastapi.py"]

requirements.txt

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

requirements_version.txt

@@ -1,48 +1,49 @@
-annotated-types==0.7.0
+annotated-types==0.7.0
-anyio==4.4.0
+anyio==4.4.0
-av==12.2.0
+av==12.3.0
-certifi==2024.7.4
+certifi==2024.8.30
-cffi==1.16.0
+cffi==1.17.1
-charset-normalizer==3.3.2
+charset-normalizer==3.3.2
-click==8.1.7
+click==8.1.7
-coloredlogs==15.0.1
+coloredlogs==15.0.1
-ctranslate2==4.3.1
+ctranslate2==4.4.0
-fastapi==0.111.0
+exceptiongroup==1.2.2
-faster-whisper==1.0.3
+fastapi==0.114.1
-filelock==3.15.4
+faster-whisper @ git+https://github.com/heimoshuiyu/faster-whisper@28a4d11a736d8cdeb4655ee5d7e4b4e7ae5ec8e0
-flatbuffers==24.3.25
+filelock==3.16.0
-fsspec==2024.6.1
+flatbuffers==24.3.25
-h11==0.14.0
+fsspec==2024.9.0
-httptools==0.6.1
+h11==0.14.0
-huggingface-hub==0.23.4
+httptools==0.6.1
-humanfriendly==10.0
+huggingface-hub==0.24.6
-idna==3.7
+humanfriendly==10.0
-mpmath==1.3.0
+idna==3.8
-numpy==1.26.4
+mpmath==1.3.0
-onnxruntime==1.18.1
+numpy==2.1.1
-OpenCC==1.1.7
+onnxruntime==1.19.2
-packaging==24.1
+OpenCC==1.1.9
-prometheus-client==0.18.0
+packaging==24.1
-prometheus-fastapi-instrumentator==7.0.0
+prometheus-fastapi-instrumentator==7.0.0
-protobuf==5.27.2
+prometheus_client==0.20.0
-pycparser==2.22
+protobuf==5.28.0
-pydantic==2.8.2
+pycparser==2.22
-pydantic_core==2.20.1
+pydantic==2.9.1
-pydub==0.25.1
+pydantic_core==2.23.3
-python-dotenv==1.0.1
+pydub==0.25.1
-python-multipart==0.0.9
+python-dotenv==1.0.1
-PyYAML==6.0.1
+python-multipart==0.0.9
-requests==2.32.3
+PyYAML==6.0.2
-sniffio==1.3.1
+requests==2.32.3
-sounddevice==0.4.7
+sniffio==1.3.1
-starlette==0.37.2
+sounddevice==0.5.0
-sympy==1.12.1
+starlette==0.38.5
-tokenizers==0.19.1
+sympy==1.13.2
-tqdm==4.66.4
+tokenizers==0.20.0
-typing_extensions==4.12.2
+tqdm==4.66.5
-urllib3==2.2.2
+typing_extensions==4.12.2
-uvicorn==0.30.1
+urllib3==2.2.2
-uvloop==0.19.0
+uvicorn==0.30.6
-watchfiles==0.22.0
+uvloop==0.20.0
-websockets==12.0
+watchfiles==0.24.0
-whisper-ctranslate2==0.4.5
+websockets==13.0.1
+whisper-ctranslate2==0.4.5

start-docker.sh

@@ -0,0 +1,10 @@
+#!/bin/bash
+
+docker run -d --name whisper-fastapi \
+    --restart unless-stopped \
+    --name whisper-fastapi \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    --gpus all \
+    -p 5000:5000 \
+    docker.io/heimoshuiyu/whisper-fastapi:latest \
+    --model large-v2

start-podman.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+
+podman run -d --name whisper-fastapi \
+    --restart unless-stopped \
+    --name whisper-fastapi \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    --device nvidia.com/gpu=all --security-opt=label=disable \
+    --gpus all \
+    -p 5000:5000 \
+    docker.io/heimoshuiyu/whisper-fastapi:latest \
+    --model large-v2

whisper_fastapi.py

@@ -39,7 +39,7 @@ app = FastAPI()
 Instrumentator().instrument(app).expose(app, endpoint="/konele/metrics")
 ccc = opencc.OpenCC("t2s.json")
 
-print("Loading model...")
+print(f"Loading model to device {args.device}...")
 transcriber = Transcribe(
     model_path=args.model,
     device=args.device,
@@ -49,7 +49,7 @@ transcriber = Transcribe(
     cache_directory=args.cache_dir,
     local_files_only=args.local_files_only,
 )
-print("Model loaded!")
+print(f"Model loaded to device {transcriber.model.model.device}")
 
 
 # allow all cors
@@ -78,7 +78,7 @@ def tsv_writer(generator: Generator[dict[str, Any], Any, None]):
     for i, segment in enumerate(generator):
         start_time = str(round(1000 * segment["start"]))
         end_time = str(round(1000 * segment["end"]))
-        text = segment["text"]
+        text = segment["text"].strip()
         yield f"{start_time}\t{end_time}\t{text}\n"
 
 
@@ -90,7 +90,7 @@ def srt_writer(generator: Generator[dict[str, Any], Any, None]):
         end_time = format_timestamp(
             segment["end"], decimal_marker=",", always_include_hours=True
         )
-        text = segment["text"]
+        text = segment["text"].strip()
         yield f"{i}\n{start_time} --> {end_time}\n{text}\n\n"
 
 
@@ -99,7 +99,7 @@ def vtt_writer(generator: Generator[dict[str, Any], Any, None]):
     for i, segment in enumerate(generator):
         start_time = format_timestamp(segment["start"])
         end_time = format_timestamp(segment["end"])
-        text = segment["text"]
+        text = segment["text"].strip()
         yield f"{start_time} --> {end_time}\n{text}\n\n"
 
 
@@ -159,6 +159,8 @@ def stream_builder(
 
 @app.websocket("/k6nele/status")
 @app.websocket("/konele/status")
+@app.websocket("/v1/k6nele/status")
+@app.websocket("/v1/konele/status")
 async def konele_status(
     websocket: WebSocket,
 ):
@@ -169,6 +171,8 @@ async def konele_status(
 
 @app.websocket("/k6nele/ws")
 @app.websocket("/konele/ws")
+@app.websocket("/v1/k6nele/ws")
+@app.websocket("/v1/konele/ws")
 async def konele_ws(
     websocket: WebSocket,
     task: Literal["transcribe", "translate"] = "transcribe",
@@ -182,15 +186,11 @@ async def konele_ws(
     # convert lang code format (eg. en-US to en)
     lang = lang.split("-")[0]
 
-    print("WebSocket client connected, lang is", lang)
-    print("content-type is", content_type)
     data = b""
     while True:
         try:
             data += await websocket.receive_bytes()
-            print("Received data:", len(data), data[-10:])
             if data[-3:] == b"EOS":
-                print("End of speech")
                 break
         except:
             break
@@ -213,17 +213,16 @@ async def konele_ws(
 
     file_obj.seek(0)
 
-    generator = stream_builder(
+    generator, info = stream_builder(
         audio=file_obj,
         task=task,
         vad_filter=vad_filter,
         language=None if lang == "und" else lang,
         initial_prompt=initial_prompt,
     )
-    result = build_json_result(generator)
+    result = build_json_result(generator, info)
 
     text = result.get("text", "")
-    print("result", text)
 
     await websocket.send_json(
         {
@@ -238,6 +237,8 @@ async def konele_ws(
 
 @app.post("/k6nele/post")
 @app.post("/konele/post")
+@app.post("/v1/k6nele/post")
+@app.post("/v1/konele/post")
 async def translateapi(
     request: Request,
     task: Literal["transcribe", "translate"] = "transcribe",
@@ -246,14 +247,12 @@ async def translateapi(
     vad_filter: bool = False,
 ):
     content_type = request.headers.get("Content-Type", "")
-    print("downloading request file", content_type)
 
     # convert lang code format (eg. en-US to en)
     lang = lang.split("-")[0]
 
     splited = [i.strip() for i in content_type.split(",") if "=" in i]
     info = {k: v for k, v in (i.split("=") for i in splited)}
-    print(info)
 
     channels = int(info.get("channels", "1"))
     rate = int(info.get("rate", "16000"))
@@ -277,17 +276,16 @@ async def translateapi(
 
     file_obj.seek(0)
 
-    generator = stream_builder(
+    generator, info = stream_builder(
         audio=file_obj,
         task=task,
         vad_filter=vad_filter,
         language=None if lang == "und" else lang,
         initial_prompt=initial_prompt,
     )
-    result = build_json_result(generator)
+    result = build_json_result(generator, info)
 
     text = result.get("text", "")
-    print("result", text)
 
     return {
         "status": 0,
@@ -327,7 +325,7 @@ async def transcription(
             media_type="text/event-stream",
         )
     elif response_format == "json":
-        return build_json_result(generator)
+        return build_json_result(generator, info)
     elif response_format == "text":
         return StreamingResponse(text_writer(generator), media_type="text/plain")
     elif response_format == "tsv":