mARTin-B78
4667144b72
- BenchRow gains a device field; Transcriber records its resolved device, so local engines report CPU or GPU and remote engines show "remote". New Device column in the results table. - Accuracy is now case-sensitive by default (capitalisation counts) so an all-lowercase transcript no longer scores 100%. Punctuation is still ignored. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
69 lines
2.6 KiB
Python
69 lines
2.6 KiB
Python
"""Local speech-to-text via faster-whisper (CTranslate2).
|
|
|
|
The model is loaded once and reused. On this arm64 host the CTranslate2 wheel is
|
|
CPU-only, so device="auto" attempts CUDA and falls back to CPU automatically.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
|
|
class Transcriber:
|
|
def __init__(
|
|
self,
|
|
model: str = "small",
|
|
device: str = "auto",
|
|
compute_type: str = "auto",
|
|
beam_size: int = 5,
|
|
):
|
|
self.beam_size = beam_size
|
|
self.device = "cpu" # resolved actual device, set by _load
|
|
self._model = self._load(model, device, compute_type)
|
|
|
|
@staticmethod
|
|
def _resolve_compute_type(device: str, requested: str) -> str:
|
|
if requested != "auto":
|
|
return requested
|
|
return "float16" if device == "cuda" else "int8"
|
|
|
|
def _load(self, model: str, device: str, compute_type: str):
|
|
from faster_whisper import WhisperModel
|
|
|
|
from .logbuffer import log
|
|
|
|
attempts: list[str]
|
|
if device == "auto":
|
|
attempts = ["cuda", "cpu"]
|
|
else:
|
|
attempts = [device]
|
|
|
|
last_err: Exception | None = None
|
|
for dev in attempts:
|
|
try:
|
|
ct = self._resolve_compute_type(dev, compute_type)
|
|
log(f"Loading Whisper '{model}' on {dev} ({ct})… (first run may download the model)")
|
|
m = WhisperModel(model, device=dev, compute_type=ct)
|
|
log(f"Whisper '{model}' ready on {dev} ({ct})")
|
|
self.device = dev
|
|
return m
|
|
except Exception as exc: # noqa: BLE001 - CUDA libs may be absent
|
|
last_err = exc
|
|
if dev != attempts[-1]:
|
|
log(f"{dev} unavailable ({exc}); trying next device")
|
|
raise RuntimeError(f"Failed to load Whisper model '{model}': {last_err}")
|
|
|
|
def transcribe(self, audio_path: Path, language: str = "", hotwords: str = "") -> str:
|
|
kwargs = dict(language=language or None, beam_size=self.beam_size, vad_filter=True)
|
|
if hotwords:
|
|
# Bias recognition toward the routing keywords so they transcribe
|
|
# reliably. Older faster-whisper builds lack `hotwords`; fall back.
|
|
kwargs["hotwords"] = hotwords
|
|
try:
|
|
segments, _info = self._model.transcribe(str(audio_path), **kwargs)
|
|
except TypeError:
|
|
kwargs.pop("hotwords", None)
|
|
segments, _info = self._model.transcribe(str(audio_path), **kwargs)
|
|
return " ".join(seg.text.strip() for seg in segments).strip()
|