115 lines
3.5 KiB
Python
115 lines
3.5 KiB
Python
import os
|
|
import sys
|
|
import tempfile
|
|
import time
|
|
import types
|
|
import importlib.machinery
|
|
from pathlib import Path
|
|
from threading import Lock
|
|
|
|
from fastapi import FastAPI, File, UploadFile
|
|
|
|
app = FastAPI(title="Emotion Museum ASR")
|
|
|
|
MODEL_NAME = os.getenv("ASR_MODEL", "/data/programs/emotion-museum/asr-service/models/paraformer-zh-onnx")
|
|
DEVICE = os.getenv("ASR_DEVICE", "cpu")
|
|
WORK_DIR = Path(os.getenv("ASR_WORK_DIR", "/tmp/emotion-museum-asr"))
|
|
WORK_DIR.mkdir(parents=True, exist_ok=True)
|
|
|
|
_model = None
|
|
_model_lock = Lock()
|
|
|
|
|
|
def get_model():
|
|
global _model
|
|
with _model_lock:
|
|
if _model is None:
|
|
# funasr-onnx imports the optional SenseVoice module from package
|
|
# __init__, which imports torch even when we only use Paraformer.
|
|
# This service intentionally runs the ONNX path without PyTorch.
|
|
if "torch" not in sys.modules:
|
|
torch_stub = types.ModuleType("torch")
|
|
torch_stub.__spec__ = importlib.machinery.ModuleSpec("torch", loader=None)
|
|
torch_stub.Tensor = type("Tensor", (), {})
|
|
sys.modules["torch"] = torch_stub
|
|
from funasr_onnx import Paraformer
|
|
|
|
_model = Paraformer(
|
|
MODEL_NAME,
|
|
batch_size=1,
|
|
device_id=-1,
|
|
quantize=True,
|
|
intra_op_num_threads=2,
|
|
)
|
|
return _model
|
|
|
|
|
|
def clean_text(text):
|
|
if isinstance(text, (list, tuple)):
|
|
text = text[0] if text else ""
|
|
if not text:
|
|
return ""
|
|
markers = ["<|zh|>", "<|en|>", "<|yue|>", "<|ja|>", "<|ko|>", "<|nospeech|>", "<|withitn|>", "<|woitn|>"]
|
|
for marker in markers:
|
|
text = text.replace(marker, "")
|
|
return text.strip()
|
|
|
|
|
|
@app.get("/health")
|
|
def health():
|
|
return {
|
|
"status": "ok",
|
|
"engine": "funasr-onnx",
|
|
"model": MODEL_NAME,
|
|
"device": DEVICE,
|
|
"loaded": _model is not None,
|
|
}
|
|
|
|
|
|
@app.post("/transcribe")
|
|
async def transcribe(file: UploadFile = File(...)):
|
|
started = time.time()
|
|
suffix = Path(file.filename or "audio.wav").suffix or ".wav"
|
|
tmp_path = None
|
|
|
|
try:
|
|
with tempfile.NamedTemporaryFile(delete=False, suffix=suffix, dir=WORK_DIR) as tmp:
|
|
tmp_path = Path(tmp.name)
|
|
while True:
|
|
chunk = await file.read(1024 * 1024)
|
|
if not chunk:
|
|
break
|
|
tmp.write(chunk)
|
|
|
|
model = get_model()
|
|
result = model([str(tmp_path)])
|
|
first = result[0] if isinstance(result, list) and result else result
|
|
text = clean_text(first.get("preds", first.get("text", "")) if isinstance(first, dict) else str(first or ""))
|
|
language = first.get("language") if isinstance(first, dict) else None
|
|
|
|
return {
|
|
"success": bool(text),
|
|
"text": text,
|
|
"language": language,
|
|
"durationMs": int((time.time() - started) * 1000),
|
|
"engine": "funasr-onnx",
|
|
"model": MODEL_NAME,
|
|
"errorMessage": None if text else "empty recognition result",
|
|
}
|
|
except Exception as exc:
|
|
return {
|
|
"success": False,
|
|
"text": "",
|
|
"language": None,
|
|
"durationMs": int((time.time() - started) * 1000),
|
|
"engine": "funasr-onnx",
|
|
"model": MODEL_NAME,
|
|
"errorMessage": str(exc),
|
|
}
|
|
finally:
|
|
if tmp_path:
|
|
try:
|
|
tmp_path.unlink(missing_ok=True)
|
|
except Exception:
|
|
pass
|