初始化

2026-08-25 17:59:42 +08:00
commit 4b7380dd9b
408 changed files with 327400 additions and 0 deletions
@@ -0,0 +1,86 @@
+"""CPU-friendly faster-whisper entry point used by the Go media worker.
+
+The process writes JSON only to stdout. Model/runtime diagnostics go to stderr so
+the Go caller never mistakes logs for transcript data.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+import tempfile
+
+
+def configure_model_cache() -> None:
+    cache_root = os.getenv("JCF_ASR_CACHE_DIR") or os.path.join(tempfile.gettempdir(), "jcf-asr-cache")
+    os.makedirs(cache_root, mode=0o700, exist_ok=True)
+    hf_home = os.getenv("HF_HOME") or os.path.join(cache_root, "huggingface")
+    os.makedirs(hf_home, mode=0o700, exist_ok=True)
+    os.environ.setdefault("XDG_CACHE_HOME", cache_root)
+    os.environ.setdefault("HF_HOME", hf_home)
+
+
+def main() -> int:
+    if len(sys.argv) < 2:
+        print(json.dumps({"error": "missing_audio_path", "segments": []}, ensure_ascii=False))
+        return 2
+
+    configure_model_cache()
+
+    from faster_whisper import WhisperModel
+
+    audio_path = sys.argv[1]
+    language = (sys.argv[2] if len(sys.argv) > 2 else "").strip() or None
+    model_name = os.getenv("FASTER_WHISPER_MODEL_PATH") or os.getenv("FASTER_WHISPER_MODEL_SIZE", "small")
+    device = os.getenv("FASTER_WHISPER_DEVICE", "cpu")
+    compute_type = os.getenv("FASTER_WHISPER_COMPUTE_TYPE", "int8")
+    model = WhisperModel(model_name, device=device, compute_type=compute_type)
+    segments, info = model.transcribe(
+        audio_path,
+        language=language,
+        beam_size=5,
+        best_of=5,
+        temperature=0.0,
+        condition_on_previous_text=True,
+        vad_filter=True,
+        vad_parameters={"min_silence_duration_ms": 700},
+        word_timestamps=True,
+    )
+    result = []
+    for segment in segments:
+        text = (segment.text or "").strip()
+        if not text:
+            continue
+        words = []
+        probabilities = []
+        for word in segment.words or []:
+            if word.start is None or word.end is None or not word.word:
+                continue
+            probability = float(word.probability) if word.probability is not None else None
+            if probability is not None:
+                probabilities.append(probability)
+            words.append({
+                "start": round(float(word.start), 3),
+                "end": round(float(word.end), 3),
+                "text": str(word.word),
+                "probability": round(probability, 4) if probability is not None else None,
+            })
+        mean_probability = sum(probabilities) / len(probabilities) if probabilities else None
+        avg_logprob = float(getattr(segment, "avg_logprob", -1.0) or -1.0)
+        no_speech = float(getattr(segment, "no_speech_prob", 0.0) or 0.0)
+        confidence = max(0.0, min(1.0, (1.0 + avg_logprob) * 0.45 + (1.0 - no_speech) * 0.2 + (mean_probability if mean_probability is not None else max(0.0, 1.0 + avg_logprob)) * 0.35))
+        result.append({
+            "start": round(float(segment.start), 3),
+            "end": round(float(segment.end), 3),
+            "text": text,
+            "words": words,
+            "confidence": round(confidence, 3),
+            "needs_review": confidence < 0.62 or (mean_probability is not None and mean_probability < 0.55),
+        })
+    print(json.dumps({"language": getattr(info, "language", language), "segments": result}, ensure_ascii=False))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())