Spaces:
Sleeping
Sleeping
Arseniy Savin commited on
Commit ·
5aea4af
1
Parent(s): 9e5239c
Implement initial setup for Whisper.cpp FastAPI application with Docker support. Added Dockerfile, main application logic, requirements, and updated README with usage instructions and environment variables.
Browse files- Dockerfile +28 -0
- README.md +25 -2
- main.py +147 -0
- requirements.txt +8 -0
Dockerfile
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# See: https://huggingface.co/docs/hub/spaces-sdks-docker
|
| 2 |
+
FROM python:3.10-slim
|
| 3 |
+
|
| 4 |
+
# Create non-root user per Spaces guidance
|
| 5 |
+
RUN useradd -m -u 1000 user
|
| 6 |
+
ENV HOME=/home/user \
|
| 7 |
+
PATH=/home/user/.local/bin:$PATH
|
| 8 |
+
WORKDIR $HOME/app
|
| 9 |
+
|
| 10 |
+
# System deps: ffmpeg for non-wav inputs
|
| 11 |
+
RUN apt-get update && apt-get install -y --no-install-recommends \
|
| 12 |
+
ffmpeg \
|
| 13 |
+
&& rm -rf /var/lib/apt/lists/*
|
| 14 |
+
|
| 15 |
+
# Python deps
|
| 16 |
+
COPY --chown=user requirements.txt ./requirements.txt
|
| 17 |
+
RUN pip install --no-cache-dir --upgrade -r requirements.txt
|
| 18 |
+
|
| 19 |
+
# App
|
| 20 |
+
COPY --chown=user . .
|
| 21 |
+
USER user
|
| 22 |
+
|
| 23 |
+
# Default env tuned for CPU Basic
|
| 24 |
+
ENV WHISPER_MODEL=tiny.en \
|
| 25 |
+
WHISPER_THREADS=4
|
| 26 |
+
|
| 27 |
+
CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860"]
|
| 28 |
+
|
README.md
CHANGED
|
@@ -4,7 +4,30 @@ emoji: 🦀
|
|
| 4 |
colorFrom: indigo
|
| 5 |
colorTo: blue
|
| 6 |
sdk: docker
|
| 7 |
-
|
| 8 |
---
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 4 |
colorFrom: indigo
|
| 5 |
colorTo: blue
|
| 6 |
sdk: docker
|
| 7 |
+
app_port: 7860
|
| 8 |
---
|
| 9 |
|
| 10 |
+
Minimal REST API for Whisper.cpp on free CPU Spaces.
|
| 11 |
+
|
| 12 |
+
Endpoints
|
| 13 |
+
|
| 14 |
+
- GET /health: liveness check
|
| 15 |
+
- POST /v1/transcribe: multipart file upload (audio/video). Returns JSON with text and per-segment metadata.
|
| 16 |
+
|
| 17 |
+
Quick use
|
| 18 |
+
|
| 19 |
+
```bash
|
| 20 |
+
curl -s -X POST \
|
| 21 |
+
-F "file=@sample.mp3" \
|
| 22 |
+
https://<your-space-subdomain>.hf.space/v1/transcribe | jq
|
| 23 |
+
```
|
| 24 |
+
|
| 25 |
+
Environment
|
| 26 |
+
|
| 27 |
+
- WHISPER_MODEL: model name (e.g. tiny.en, base.en, small). Default tiny.en.
|
| 28 |
+
- WHISPER_THREADS: number of CPU threads to use. Default 4.
|
| 29 |
+
|
| 30 |
+
Notes
|
| 31 |
+
|
| 32 |
+
- Uses pywhispercpp wheels (no compile) + ffmpeg to handle non-WAV inputs.
|
| 33 |
+
- Weights cache in ephemeral storage. For persistence, add a storage upgrade and set HF_HOME=/data/.huggingface.
|
main.py
ADDED
|
@@ -0,0 +1,147 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
from typing import Any, Dict, List, Optional
|
| 3 |
+
|
| 4 |
+
import numpy as np # type: ignore[import]
|
| 5 |
+
import orjson # type: ignore[import]
|
| 6 |
+
from fastapi import ( # type: ignore[import]
|
| 7 |
+
FastAPI,
|
| 8 |
+
File,
|
| 9 |
+
UploadFile,
|
| 10 |
+
HTTPException,
|
| 11 |
+
)
|
| 12 |
+
from fastapi.responses import ( # type: ignore[import]
|
| 13 |
+
JSONResponse,
|
| 14 |
+
)
|
| 15 |
+
from pydantic import BaseModel # type: ignore[import]
|
| 16 |
+
import ffmpeg # type: ignore[import]
|
| 17 |
+
|
| 18 |
+
from pywhispercpp.model import Model # type: ignore[import]
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
def json_dumps(v: Any, *, default: Optional[Any] = None) -> str:
|
| 22 |
+
return orjson.dumps(v, default=default).decode()
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class Segment(BaseModel):
|
| 26 |
+
id: int
|
| 27 |
+
start: float
|
| 28 |
+
end: float
|
| 29 |
+
text: str
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
class TranscribeResponse(BaseModel):
|
| 33 |
+
model: str
|
| 34 |
+
language: Optional[str] = None
|
| 35 |
+
duration: Optional[float] = None
|
| 36 |
+
text: str
|
| 37 |
+
segments: List[Segment]
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
app = FastAPI(title="Whisper.cpp FastAPI", default_response_class=JSONResponse)
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
def load_model() -> Model:
|
| 44 |
+
model_name = os.getenv("WHISPER_MODEL", "tiny.en")
|
| 45 |
+
return Model(model_name)
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
WHISPER_MODEL_INSTANCE: Optional[Model] = None
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
@app.on_event("startup")
|
| 52 |
+
def _startup() -> None:
|
| 53 |
+
global WHISPER_MODEL_INSTANCE
|
| 54 |
+
WHISPER_MODEL_INSTANCE = load_model()
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
@app.get("/health")
|
| 58 |
+
def health() -> Dict[str, str]:
|
| 59 |
+
return {"status": "ok"}
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
def _decode_audio_to_wav_bytes(
|
| 63 |
+
file_bytes: bytes, sample_rate: int = 16000
|
| 64 |
+
) -> bytes:
|
| 65 |
+
try:
|
| 66 |
+
out, _ = (
|
| 67 |
+
ffmpeg.input("pipe:0")
|
| 68 |
+
.output(
|
| 69 |
+
"pipe:1",
|
| 70 |
+
format="s16le",
|
| 71 |
+
acodec="pcm_s16le",
|
| 72 |
+
ac=1,
|
| 73 |
+
ar=str(sample_rate),
|
| 74 |
+
)
|
| 75 |
+
.run(
|
| 76 |
+
input=file_bytes,
|
| 77 |
+
cmd=["ffmpeg", "-nostdin"],
|
| 78 |
+
capture_stdout=True,
|
| 79 |
+
capture_stderr=True,
|
| 80 |
+
)
|
| 81 |
+
)
|
| 82 |
+
return out
|
| 83 |
+
except ffmpeg.Error as e:
|
| 84 |
+
detail = f"ffmpeg decode error: {e.stderr.decode()}"
|
| 85 |
+
raise HTTPException(status_code=400, detail=detail) from e
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
def _bytes_to_float32_mono(arr_bytes: bytes) -> np.ndarray:
|
| 89 |
+
int16 = np.frombuffer(arr_bytes, dtype=np.int16)
|
| 90 |
+
return (int16.astype(np.float32) / 32768.0).flatten()
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
@app.post("/v1/transcribe", response_model=TranscribeResponse)
|
| 94 |
+
async def transcribe(file: UploadFile = File(...)) -> TranscribeResponse:
|
| 95 |
+
if WHISPER_MODEL_INSTANCE is None:
|
| 96 |
+
raise HTTPException(status_code=503, detail="model not loaded")
|
| 97 |
+
|
| 98 |
+
file_bytes = await file.read()
|
| 99 |
+
if not file_bytes:
|
| 100 |
+
raise HTTPException(status_code=400, detail="empty file")
|
| 101 |
+
|
| 102 |
+
# Decode with ffmpeg to 16k mono PCM
|
| 103 |
+
pcm_bytes = _decode_audio_to_wav_bytes(file_bytes, sample_rate=16000)
|
| 104 |
+
audio = _bytes_to_float32_mono(pcm_bytes)
|
| 105 |
+
|
| 106 |
+
# Threads hint
|
| 107 |
+
threads_env = os.getenv("WHISPER_THREADS")
|
| 108 |
+
try:
|
| 109 |
+
n_threads = int(threads_env) if threads_env else None
|
| 110 |
+
except ValueError:
|
| 111 |
+
n_threads = None
|
| 112 |
+
|
| 113 |
+
# Run transcription
|
| 114 |
+
segments = WHISPER_MODEL_INSTANCE.transcribe(
|
| 115 |
+
audio, num_proc=n_threads or 1
|
| 116 |
+
)
|
| 117 |
+
|
| 118 |
+
seg_models: List[Segment] = []
|
| 119 |
+
full_text_parts: List[str] = []
|
| 120 |
+
for idx, seg in enumerate(segments):
|
| 121 |
+
# pywhispercpp returns objects with .text/.t0/.t1 timestamps in 0.01s
|
| 122 |
+
# units; normalize to seconds if present
|
| 123 |
+
start_s = getattr(seg, "start", None)
|
| 124 |
+
end_s = getattr(seg, "end", None)
|
| 125 |
+
text = getattr(seg, "text", "")
|
| 126 |
+
# Some versions expose t0/t1 in 0.01s ticks
|
| 127 |
+
if start_s is None and hasattr(seg, "t0"):
|
| 128 |
+
start_s = float(seg.t0) / 100.0
|
| 129 |
+
if end_s is None and hasattr(seg, "t1"):
|
| 130 |
+
end_s = float(seg.t1) / 100.0
|
| 131 |
+
seg_models.append(
|
| 132 |
+
Segment(
|
| 133 |
+
id=idx,
|
| 134 |
+
start=float(start_s or 0.0),
|
| 135 |
+
end=float(end_s or 0.0),
|
| 136 |
+
text=text,
|
| 137 |
+
)
|
| 138 |
+
)
|
| 139 |
+
if text:
|
| 140 |
+
full_text_parts.append(text)
|
| 141 |
+
|
| 142 |
+
resp = TranscribeResponse(
|
| 143 |
+
model=os.getenv("WHISPER_MODEL", "tiny.en"),
|
| 144 |
+
text="".join(full_text_parts).strip(),
|
| 145 |
+
segments=seg_models,
|
| 146 |
+
)
|
| 147 |
+
return resp
|
requirements.txt
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
fastapi==0.110.0
|
| 2 |
+
uvicorn[standard]==0.27.1
|
| 3 |
+
python-multipart==0.0.9
|
| 4 |
+
pywhispercpp==1.3.3
|
| 5 |
+
ffmpeg-python==0.2.0
|
| 6 |
+
numpy==1.26.4
|
| 7 |
+
pydantic==2.6.3
|
| 8 |
+
orjson==3.9.15
|