app.py

import os
import tempfile
import pandas as pd
import gradio as gr
from pydub import AudioSegment
from faster_whisper import WhisperModel
from pyannote.audio import Pipeline as DiarizationPipeline

# Initialisation des modèles
whisper_model = WhisperModel("large-v2", device="cpu", compute_type="int8")
token = os.getenv("HF_TOKEN")  # Ajoute cette variable dans les secrets de ton Space si besoin
diari_pipeline = DiarizationPipeline.from_pretrained(
    "pyannote/speaker-diarization-3.1",
    use_auth_token=token
)

# Pipeline de traitement :
# .mp3
#  ↓ (converti .wav)
# .wav
#  ↓
# faster-whisper → segments (texte + timestamps)
#  ↓
# pyannote-audio → diarisation (segments + speaker X)
#  ↓
# Fusion des deux → transcription enrichie avec speaker + timestamp

def convert_mp3_to_wav(mp3_path):
    wav_path = tempfile.mktemp(suffix=".wav")
    audio = AudioSegment.from_file(mp3_path, format="mp3")
    audio = audio.set_channels(1).set_frame_rate(16000)
    audio.export(wav_path, format="wav")
    return wav_path

def transcribe_and_diarize(audio_file):
    wav_path = convert_mp3_to_wav(audio_file)

    # Transcription
    segments, _ = whisper_model.transcribe(wav_path, language="fr", beam_size=5)

    # Diarisation
    diarization = diari_pipeline(wav_path)
    speakers = []
    for turn, _, speaker in diarization.itertracks(yield_label=True):
        speakers.append({
            "start": turn.start,
            "end": turn.end,
            "speaker": speaker
        })

    # Fusion transcription + speaker
    final_output = []
    for seg in segments:
        seg_start = seg.start
        seg_end = seg.end
        text = seg.text.strip()
        speaker = "Inconnu"
        for s in speakers:
            if s["start"] <= seg_start <= s["end"]:
                speaker = s["speaker"]
                break
        final_output.append({
            "start": seg_start,
            "end": seg_end,
            "speaker": speaker,
            "text": text
        })

    df = pd.DataFrame(final_output)
    txt_lines = [f"[{row['start']:.2f}s - {row['end']:.2f}s] {row['speaker']} : {row['text']}" for _, row in df.iterrows()]
    txt_output = "\n".join(txt_lines)

    txt_path = tempfile.mktemp(suffix=".txt")
    with open(txt_path, "w", encoding="utf-8") as f:
        f.write(txt_output)

    csv_path = tempfile.mktemp(suffix=".csv")
    df.to_csv(csv_path, index=False)

    return txt_output, csv_path, txt_path

# Interface Gradio
gr.Interface(
    fn=transcribe_and_diarize,
    inputs=gr.Audio(type="filepath", label="Fichier audio MP3"),
    outputs=[
        gr.Textbox(label="Transcription avec locuteurs"),
        gr.File(label="Télécharger le CSV"),
        gr.File(label="Télécharger le TXT")
    ],
    title="Transcription + Diarisation (FR)",
    description="Charge un fichier MP3. Transcription FR + séparation des locuteurs + export CSV/TXT."
).launch()