File size: 2,292 Bytes
d797bdc 4e3de1b d797bdc 4e3de1b d797bdc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | """
Whisper STT wrapper — ROCm-compatible.
Two modes:
1. Local Whisper (runs on the same MI300X, via openai-whisper or faster-whisper)
2. Remote endpoint (OpenAI-compatible /v1/audio/transcriptions)
Set STT_MODE env var to "local" or "remote". Defaults to "remote" in HF Space.
"""
import os
from pathlib import Path
STT_MODE = os.getenv("STT_MODE", "remote")
STT_URL = os.getenv("STT_URL", "")
WHISPER_SIZE = os.getenv("WHISPER_SIZE", "base") # tiny/base/small/medium
# ---------------------------------------------------------------------------
# Remote path (OpenAI-compatible endpoint — used in HF Space)
# ---------------------------------------------------------------------------
def _transcribe_remote(audio_path: str) -> str:
import httpx
url = f"{STT_URL.rstrip('/')}/v1/audio/transcriptions"
with open(audio_path, "rb") as f:
r = httpx.post(
url,
files={"file": (os.path.basename(audio_path), f, "audio/wav")},
data={"model": "whisper-1"},
timeout=30.0,
)
r.raise_for_status()
return r.json()["text"].strip()
# ---------------------------------------------------------------------------
# Local path (runs on MI300X with ROCm)
# ---------------------------------------------------------------------------
_whisper_model = None
def _load_whisper():
global _whisper_model
if _whisper_model is None:
import whisper
_whisper_model = whisper.load_model(WHISPER_SIZE)
return _whisper_model
def _transcribe_local(audio_path: str) -> str:
model = _load_whisper()
result = model.transcribe(audio_path, fp16=True, language="en")
return result["text"].strip()
# ---------------------------------------------------------------------------
# Public API
# ---------------------------------------------------------------------------
def transcribe(audio_path: str) -> str:
"""Transcribe audio file to text. Returns empty string on failure."""
if not audio_path or not Path(audio_path).exists():
return ""
try:
if STT_MODE == "local":
return _transcribe_local(audio_path)
return _transcribe_remote(audio_path)
except Exception as e:
print(f"[STT] Error: {e}")
return ""
|