#!/usr/bin/env python3
"""ASR local para o Italiano Vivo via faster-whisper.

O proxy LiteLLM (10.8.0.1:4000) nao aceita partes de audio no chat/completions
(qwen omni devolve 400 "Invalid video file" / "provided URL does not appear to be valid"),
entao a transcricao roda aqui e o server Node repassa so o texto para o juiz LLM.

Uso:
    python asr_whisper.py <audio_path> [--model small] [--json]

Imprime o texto transcrito (ou JSON com segments/timing com --json).
Idempotente: baixa o modelo uma vez e usa cache local do HuggingFace.
"""
import argparse
import json
import sys


def main() -> int:
    ap = argparse.ArgumentParser(description="Transcricao IT/PT com faster-whisper")
    ap.add_argument("audio", help="caminho do arquivo de audio (mp3/wav/webm)")
    ap.add_argument("--model", default="small", help="tiny|base|small|medium (default small)")
    ap.add_argument("--lang", default=None, help="forcar idioma, ex.: it")
    ap.add_argument("--json", action="store_true", dest="as_json")
    args = ap.parse_args()

    from faster_whisper import WhisperModel

    model = WhisperModel(args.model, device="cpu", compute_type="int8")
    segments, info = model.transcribe(
        args.audio, language=args.lang, beam_size=5, vad_filter=True
    )

    parts = []
    seg_list = []
    for s in segments:
        parts.append(s.text.strip())
        seg_list.append({"start": round(s.start, 2), "end": round(s.end, 2), "text": s.text.strip()})

    text = " ".join(p for p in parts if p).strip()

    if args.as_json:
        print(json.dumps({
            "text": text,
            "language": info.language,
            "language_probability": round(float(info.language_probability), 3),
            "segments": seg_list,
        }, ensure_ascii=False))
    else:
        print(text)
    return 0 if text else 2


if __name__ == "__main__":
    sys.exit(main())