from pathlib import Path


class ProviderDeDiarizacaoHuggingFace:
    """Identifica intervalos de falas por participante em áudio local."""

    def __init__(self, modelo: str) -> None:
        from pyannote.audio import Pipeline
        self.pipeline = Pipeline.from_pretrained(modelo)

    def analisar(self, arquivo: str | Path) -> list[tuple[float, float, str]]:
        import soundfile as sf
        import torch
        audio, taxa = sf.read(str(arquivo), dtype="float32")
        if getattr(audio, "ndim", 1) > 1:
            audio = audio.mean(axis=1)
        saida = self.pipeline({"waveform": torch.from_numpy(audio).unsqueeze(0),
                               "sample_rate": taxa})
        diarizacao = getattr(saida, "exclusive_speaker_diarization", saida)
        return [(float(turno.start), float(turno.end), str(falante))
                for turno, _, falante in diarizacao.itertracks(yield_label=True)]