Spaces:
Runtime error
Runtime error
mathisescriva commited on
Commit ·
e6e14b8
1
Parent(s): 704669a
Initial commit: STT + Diarization pipeline unifié
Browse files- README.md +46 -16
- app.py +126 -90
- processing.py +360 -0
- requirements.txt +8 -0
README.md
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
---
|
| 2 |
-
title: Gilbert -
|
| 3 |
emoji: 🎤
|
| 4 |
colorFrom: blue
|
| 5 |
colorTo: purple
|
|
@@ -10,32 +10,62 @@ pinned: false
|
|
| 10 |
license: mit
|
| 11 |
---
|
| 12 |
|
| 13 |
-
# Gilbert -
|
| 14 |
|
| 15 |
-
|
| 16 |
|
| 17 |
## Fonctionnalités
|
| 18 |
|
| 19 |
-
- 🎤 Diarisation de locuteurs
|
| 20 |
-
-
|
| 21 |
-
-
|
| 22 |
-
-
|
| 23 |
|
| 24 |
-
## Modèles
|
| 25 |
|
| 26 |
-
|
| 27 |
-
- `pyannote/speaker-diarization-community-1`
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
## Utilisation
|
| 30 |
|
| 31 |
-
1. Uploadez un fichier audio (WAV, MP3, M4A)
|
| 32 |
-
2. Configurez les paramètres (optionnel)
|
| 33 |
-
3. Cliquez sur "
|
| 34 |
-
4. Téléchargez
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
|
| 36 |
## Configuration
|
| 37 |
|
| 38 |
-
Pour utiliser cette Space, vous devez avoir un token Hugging Face avec accès aux modèles pyannote.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
|
| 40 |
-
|
| 41 |
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Gilbert - STT + Diarization
|
| 3 |
emoji: 🎤
|
| 4 |
colorFrom: blue
|
| 5 |
colorTo: purple
|
|
|
|
| 10 |
license: mit
|
| 11 |
---
|
| 12 |
|
| 13 |
+
# Gilbert - STT + Diarization
|
| 14 |
|
| 15 |
+
Pipeline complet de transcription (STT) et diarisation de locuteurs avec sortie formatée.
|
| 16 |
|
| 17 |
## Fonctionnalités
|
| 18 |
|
| 19 |
+
- 🎤 **Diarisation de locuteurs** avec pyannote.audio
|
| 20 |
+
- 📝 **Transcription** avec Whisper Large V3 French (fine-tuné pour le français)
|
| 21 |
+
- 🔗 **Combinaison automatique** pour une sortie formatée: "Speaker A : texte"
|
| 22 |
+
- 📊 **Statistiques détaillées** par locuteur
|
| 23 |
|
| 24 |
+
## Modèles utilisés
|
| 25 |
|
| 26 |
+
### Diarization
|
| 27 |
+
- `pyannote/speaker-diarization-community-1` (par défaut, meilleures performances)
|
| 28 |
+
- `pyannote/speaker-diarization-3.1` (fallback)
|
| 29 |
+
|
| 30 |
+
### Speech-to-Text (STT)
|
| 31 |
+
- `bofenghuang/whisper-large-v3-french` (Whisper Large V3 fine-tuné pour le français)
|
| 32 |
+
- Meilleures performances sur le français que Whisper standard
|
| 33 |
+
- Support de la casse, ponctuation et nombres
|
| 34 |
|
| 35 |
## Utilisation
|
| 36 |
|
| 37 |
+
1. Uploadez un fichier audio (WAV, MP3, M4A, FLAC)
|
| 38 |
+
2. Configurez les paramètres de diarisation (optionnel)
|
| 39 |
+
3. Cliquez sur "Traiter"
|
| 40 |
+
4. Téléchargez la transcription avec identification des locuteurs
|
| 41 |
+
|
| 42 |
+
## Format de sortie
|
| 43 |
+
|
| 44 |
+
La sortie est au format :
|
| 45 |
+
```
|
| 46 |
+
Speaker A : texte du locuteur A
|
| 47 |
+
|
| 48 |
+
Speaker B : texte du locuteur B
|
| 49 |
+
```
|
| 50 |
|
| 51 |
## Configuration
|
| 52 |
|
| 53 |
+
Pour utiliser cette Space, vous devez avoir un token Hugging Face avec accès aux modèles pyannote et Whisper.
|
| 54 |
+
|
| 55 |
+
Configurez-le dans les secrets de la Space avec: `HF_TOKEN="votre_token"`
|
| 56 |
+
|
| 57 |
+
## Exemple de sortie
|
| 58 |
+
|
| 59 |
+
```
|
| 60 |
+
Speaker A : Bonjour, comment allez-vous aujourd'hui ?
|
| 61 |
+
|
| 62 |
+
Speaker B : Très bien merci, et vous ?
|
| 63 |
+
|
| 64 |
+
Speaker A : Parfait, je suis ravi de vous rencontrer.
|
| 65 |
+
```
|
| 66 |
|
| 67 |
+
## Performance
|
| 68 |
|
| 69 |
+
- **Temps de traitement**: ~1.5x la durée de l'audio (sur CPU)
|
| 70 |
+
- **Précision**: Optimisée pour le français avec le modèle fine-tuné
|
| 71 |
+
- **Formats supportés**: WAV, MP3, M4A, FLAC, OGG
|
app.py
CHANGED
|
@@ -4,90 +4,147 @@ import tempfile
|
|
| 4 |
from pathlib import Path
|
| 5 |
import sys
|
| 6 |
|
| 7 |
-
#
|
| 8 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
| 10 |
-
from diarization_pyannote_demo import run_pyannote_diarization, write_rttm, write_json
|
| 11 |
|
| 12 |
-
def
|
| 13 |
-
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
if audio_file is None:
|
| 16 |
return None, "❌ Veuillez uploader un fichier audio"
|
| 17 |
|
| 18 |
try:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
# Créer un répertoire temporaire pour les résultats
|
| 20 |
with tempfile.TemporaryDirectory() as tmpdir:
|
| 21 |
-
#
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
show_progress=False
|
| 31 |
-
)
|
| 32 |
|
| 33 |
-
#
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
-
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
-
#
|
| 42 |
-
|
| 43 |
-
rttm_content = f.read()
|
| 44 |
|
| 45 |
-
|
| 46 |
-
|
|
|
|
|
|
|
| 47 |
|
| 48 |
# Créer un résumé
|
| 49 |
-
summary = f"""
|
| 50 |
-
# Résultats de diarisation
|
| 51 |
-
|
| 52 |
-
**Fichier:** {Path(audio_file.name).name}
|
| 53 |
-
**Modèle:** {model_name}
|
| 54 |
-
**Locuteurs détectés:** {result['num_speakers']}
|
| 55 |
-
**Segments:** {len(result['segments'])}
|
| 56 |
-
**Durée totale:** {result.get('duration', 0):.2f} secondes
|
| 57 |
-
|
| 58 |
-
## Statistiques par locuteur
|
| 59 |
-
"""
|
| 60 |
from collections import defaultdict
|
| 61 |
-
speaker_stats = defaultdict(lambda: {"total_duration": 0.0, "num_segments": 0})
|
| 62 |
-
for seg in
|
| 63 |
speaker = seg["speaker"]
|
| 64 |
duration = seg["end"] - seg["start"]
|
| 65 |
speaker_stats[speaker]["total_duration"] += duration
|
| 66 |
speaker_stats[speaker]["num_segments"] += 1
|
|
|
|
| 67 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
for speaker, stats in sorted(speaker_stats.items()):
|
|
|
|
|
|
|
| 69 |
avg_duration = stats["total_duration"] / stats["num_segments"] if stats["num_segments"] > 0 else 0
|
| 70 |
-
summary += f"\n- **{
|
| 71 |
|
| 72 |
-
return
|
| 73 |
|
| 74 |
except Exception as e:
|
| 75 |
import traceback
|
| 76 |
-
|
| 77 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 78 |
|
| 79 |
# Interface Gradio
|
| 80 |
-
with gr.Blocks(title="Gilbert -
|
| 81 |
gr.Markdown("""
|
| 82 |
-
# 🎤 Gilbert -
|
|
|
|
|
|
|
| 83 |
|
| 84 |
-
|
|
|
|
|
|
|
|
|
|
| 85 |
|
| 86 |
**Instructions:**
|
| 87 |
1. Uploadez un fichier audio (WAV, MP3, M4A)
|
| 88 |
-
2. Configurez les paramètres (optionnel)
|
| 89 |
-
3. Cliquez sur "
|
| 90 |
-
4. Téléchargez
|
| 91 |
""")
|
| 92 |
|
| 93 |
with gr.Row():
|
|
@@ -97,60 +154,39 @@ with gr.Blocks(title="Gilbert - Diarisation pyannote") as demo:
|
|
| 97 |
type="filepath"
|
| 98 |
)
|
| 99 |
|
| 100 |
-
|
| 101 |
choices=[
|
| 102 |
-
"pyannote/speaker-diarization-3.1",
|
| 103 |
"pyannote/speaker-diarization-community-1",
|
|
|
|
| 104 |
],
|
| 105 |
-
value="pyannote/speaker-diarization-
|
| 106 |
-
label="Modèle
|
| 107 |
)
|
| 108 |
|
| 109 |
-
|
| 110 |
-
num_speakers = gr.Number(
|
| 111 |
-
label="Nombre exact de locuteurs",
|
| 112 |
-
value=0,
|
| 113 |
-
minimum=0,
|
| 114 |
-
info="0 = auto-détection"
|
| 115 |
-
)
|
| 116 |
-
min_speakers = gr.Number(
|
| 117 |
-
label="Min locuteurs",
|
| 118 |
-
value=0,
|
| 119 |
-
minimum=0,
|
| 120 |
-
info="0 = pas de limite"
|
| 121 |
-
)
|
| 122 |
-
max_speakers = gr.Number(
|
| 123 |
-
label="Max locuteurs",
|
| 124 |
-
value=0,
|
| 125 |
-
minimum=0,
|
| 126 |
-
info="0 = pas de limite"
|
| 127 |
-
)
|
| 128 |
-
|
| 129 |
-
use_exclusive = gr.Checkbox(
|
| 130 |
-
label="Exclusive speaker diarization",
|
| 131 |
-
value=False,
|
| 132 |
-
info="Simplifie la réconciliation avec transcription"
|
| 133 |
-
)
|
| 134 |
-
|
| 135 |
-
diarize_btn = gr.Button("🎯 Diariser", variant="primary")
|
| 136 |
|
| 137 |
with gr.Column():
|
| 138 |
summary_output = gr.Markdown(label="Résumé")
|
| 139 |
-
|
| 140 |
-
|
|
|
|
|
|
|
| 141 |
|
| 142 |
-
|
| 143 |
-
fn=
|
| 144 |
-
inputs=[audio_input,
|
| 145 |
-
outputs=[
|
| 146 |
)
|
| 147 |
|
| 148 |
gr.Markdown("""
|
| 149 |
---
|
| 150 |
-
**Note:** Vous devez avoir un token Hugging Face configuré avec accès aux modèles pyannote.
|
| 151 |
-
Configurez-le avec: `
|
|
|
|
|
|
|
|
|
|
|
|
|
| 152 |
""")
|
| 153 |
|
| 154 |
if __name__ == "__main__":
|
| 155 |
demo.launch()
|
| 156 |
-
|
|
|
|
| 4 |
from pathlib import Path
|
| 5 |
import sys
|
| 6 |
|
| 7 |
+
# Importer le module de traitement
|
| 8 |
+
from processing import (
|
| 9 |
+
run_diarization,
|
| 10 |
+
run_transcription,
|
| 11 |
+
combine_diarization_transcription,
|
| 12 |
+
format_output
|
| 13 |
+
)
|
| 14 |
|
|
|
|
| 15 |
|
| 16 |
+
def process_audio_stt_diarization(
|
| 17 |
+
audio_file,
|
| 18 |
+
diarization_model
|
| 19 |
+
):
|
| 20 |
+
"""Interface Gradio pour STT + Diarization combinés."""
|
| 21 |
|
| 22 |
if audio_file is None:
|
| 23 |
return None, "❌ Veuillez uploader un fichier audio"
|
| 24 |
|
| 25 |
try:
|
| 26 |
+
# Gérer le chemin du fichier audio
|
| 27 |
+
if isinstance(audio_file, tuple):
|
| 28 |
+
audio_path = audio_file[1] if len(audio_file) > 1 else audio_file[0]
|
| 29 |
+
elif isinstance(audio_file, str):
|
| 30 |
+
audio_path = audio_file
|
| 31 |
+
elif hasattr(audio_file, 'name'):
|
| 32 |
+
audio_path = audio_file.name
|
| 33 |
+
else:
|
| 34 |
+
audio_path = str(audio_file)
|
| 35 |
+
|
| 36 |
+
if not os.path.exists(audio_path):
|
| 37 |
+
return None, f"❌ Fichier audio introuvable: {audio_path}"
|
| 38 |
+
|
| 39 |
+
# Récupérer le token HF
|
| 40 |
+
hf_token = os.environ.get("HF_TOKEN")
|
| 41 |
+
if not hf_token:
|
| 42 |
+
return None, "❌ Token Hugging Face non configuré (HF_TOKEN)"
|
| 43 |
+
|
| 44 |
# Créer un répertoire temporaire pour les résultats
|
| 45 |
with tempfile.TemporaryDirectory() as tmpdir:
|
| 46 |
+
# Étape 1: Diarisation
|
| 47 |
+
try:
|
| 48 |
+
diarization_segments = run_diarization(
|
| 49 |
+
audio_path,
|
| 50 |
+
hf_token,
|
| 51 |
+
model_name=diarization_model
|
| 52 |
+
)
|
| 53 |
+
except Exception as e:
|
| 54 |
+
return None, f"❌ Erreur lors de la diarisation: {str(e)}"
|
|
|
|
|
|
|
| 55 |
|
| 56 |
+
# Étape 2: Transcription
|
| 57 |
+
try:
|
| 58 |
+
transcription_segments = run_transcription(
|
| 59 |
+
audio_path,
|
| 60 |
+
hf_token=hf_token
|
| 61 |
+
)
|
| 62 |
+
except Exception as e:
|
| 63 |
+
return None, f"❌ Erreur lors de la transcription: {str(e)}"
|
| 64 |
|
| 65 |
+
# Étape 3: Combinaison
|
| 66 |
+
try:
|
| 67 |
+
combined = combine_diarization_transcription(
|
| 68 |
+
diarization_segments,
|
| 69 |
+
transcription_segments
|
| 70 |
+
)
|
| 71 |
+
except Exception as e:
|
| 72 |
+
return None, f"❌ Erreur lors de la combinaison: {str(e)}"
|
| 73 |
|
| 74 |
+
# Étape 4: Formatage
|
| 75 |
+
formatted_text = format_output(combined)
|
|
|
|
| 76 |
|
| 77 |
+
# Sauvegarder dans un fichier temporaire
|
| 78 |
+
output_file = os.path.join(tmpdir, "transcription.txt")
|
| 79 |
+
with open(output_file, 'w', encoding='utf-8') as f:
|
| 80 |
+
f.write(formatted_text)
|
| 81 |
|
| 82 |
# Créer un résumé
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
from collections import defaultdict
|
| 84 |
+
speaker_stats = defaultdict(lambda: {"total_duration": 0.0, "num_segments": 0, "text_length": 0})
|
| 85 |
+
for seg in combined:
|
| 86 |
speaker = seg["speaker"]
|
| 87 |
duration = seg["end"] - seg["start"]
|
| 88 |
speaker_stats[speaker]["total_duration"] += duration
|
| 89 |
speaker_stats[speaker]["num_segments"] += 1
|
| 90 |
+
speaker_stats[speaker]["text_length"] += len(seg["text"])
|
| 91 |
|
| 92 |
+
summary = f"""
|
| 93 |
+
# Résultats STT + Diarization
|
| 94 |
+
|
| 95 |
+
**Fichier:** {Path(audio_path).name}
|
| 96 |
+
**Modèle diarization:** {diarization_model}
|
| 97 |
+
**Modèle STT:** bofenghuang/whisper-large-v3-french
|
| 98 |
+
**Locuteurs détectés:** {len(speaker_stats)}
|
| 99 |
+
**Segments combinés:** {len(combined)}
|
| 100 |
+
|
| 101 |
+
## Statistiques par locuteur
|
| 102 |
+
"""
|
| 103 |
for speaker, stats in sorted(speaker_stats.items()):
|
| 104 |
+
speaker_num = int(speaker.replace("SPEAKER_", ""))
|
| 105 |
+
speaker_name = f"Speaker {chr(65 + speaker_num)}"
|
| 106 |
avg_duration = stats["total_duration"] / stats["num_segments"] if stats["num_segments"] > 0 else 0
|
| 107 |
+
summary += f"\n- **{speaker_name}**: {stats['num_segments']} segments, {stats['total_duration']:.2f}s total, {avg_duration:.2f}s moyenne/segment, {stats['text_length']} caractères"
|
| 108 |
|
| 109 |
+
return output_file, summary
|
| 110 |
|
| 111 |
except Exception as e:
|
| 112 |
import traceback
|
| 113 |
+
error_details = traceback.format_exc()
|
| 114 |
+
error_msg = f"""❌ **Erreur lors du traitement**
|
| 115 |
+
|
| 116 |
+
**Message:** {str(e)}
|
| 117 |
+
|
| 118 |
+
**Détails techniques:**
|
| 119 |
+
```
|
| 120 |
+
{error_details}
|
| 121 |
+
```
|
| 122 |
+
|
| 123 |
+
**Solutions possibles:**
|
| 124 |
+
- Vérifiez que le fichier audio est valide
|
| 125 |
+
- Assurez-vous que le token HF_TOKEN est configuré dans les secrets de la Space
|
| 126 |
+
- Réessayez avec un fichier audio plus court
|
| 127 |
+
"""
|
| 128 |
+
return None, error_msg
|
| 129 |
+
|
| 130 |
|
| 131 |
# Interface Gradio
|
| 132 |
+
with gr.Blocks(title="Gilbert - STT + Diarization") as demo:
|
| 133 |
gr.Markdown("""
|
| 134 |
+
# 🎤 Gilbert - STT + Diarization
|
| 135 |
+
|
| 136 |
+
Pipeline complet de transcription (STT) et diarisation de locuteurs.
|
| 137 |
|
| 138 |
+
**Fonctionnalités:**
|
| 139 |
+
- 🎤 Diarisation de locuteurs avec pyannote.audio
|
| 140 |
+
- 📝 Transcription avec Whisper Large V3 French (fine-tuné pour le français)
|
| 141 |
+
- 🔗 Combinaison automatique pour une sortie formatée: "Speaker A : texte"
|
| 142 |
|
| 143 |
**Instructions:**
|
| 144 |
1. Uploadez un fichier audio (WAV, MP3, M4A)
|
| 145 |
+
2. Configurez les paramètres de diarisation (optionnel)
|
| 146 |
+
3. Cliquez sur "Traiter"
|
| 147 |
+
4. Téléchargez la transcription avec identification des locuteurs
|
| 148 |
""")
|
| 149 |
|
| 150 |
with gr.Row():
|
|
|
|
| 154 |
type="filepath"
|
| 155 |
)
|
| 156 |
|
| 157 |
+
diarization_model = gr.Dropdown(
|
| 158 |
choices=[
|
|
|
|
| 159 |
"pyannote/speaker-diarization-community-1",
|
| 160 |
+
"pyannote/speaker-diarization-3.1",
|
| 161 |
],
|
| 162 |
+
value="pyannote/speaker-diarization-community-1",
|
| 163 |
+
label="Modèle de diarisation"
|
| 164 |
)
|
| 165 |
|
| 166 |
+
process_btn = gr.Button("🚀 Traiter", variant="primary")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 167 |
|
| 168 |
with gr.Column():
|
| 169 |
summary_output = gr.Markdown(label="Résumé")
|
| 170 |
+
transcription_output = gr.File(
|
| 171 |
+
label="Transcription (format: Speaker A : texte)",
|
| 172 |
+
type="filepath"
|
| 173 |
+
)
|
| 174 |
|
| 175 |
+
process_btn.click(
|
| 176 |
+
fn=process_audio_stt_diarization,
|
| 177 |
+
inputs=[audio_input, diarization_model],
|
| 178 |
+
outputs=[transcription_output, summary_output]
|
| 179 |
)
|
| 180 |
|
| 181 |
gr.Markdown("""
|
| 182 |
---
|
| 183 |
+
**Note:** Vous devez avoir un token Hugging Face configuré avec accès aux modèles pyannote et Whisper.
|
| 184 |
+
Configurez-le dans les secrets de la Space avec: `HF_TOKEN="votre_token"`
|
| 185 |
+
|
| 186 |
+
**Modèles utilisés:**
|
| 187 |
+
- **Diarization**: pyannote/speaker-diarization-community-1 (ou 3.1)
|
| 188 |
+
- **STT**: bofenghuang/whisper-large-v3-french (Whisper Large V3 fine-tuné pour le français)
|
| 189 |
""")
|
| 190 |
|
| 191 |
if __name__ == "__main__":
|
| 192 |
demo.launch()
|
|
|
processing.py
ADDED
|
@@ -0,0 +1,360 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Module de traitement unifié pour STT + Diarization.
|
| 4 |
+
Utilisé par le Space Gradio.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import os
|
| 8 |
+
import sys
|
| 9 |
+
from pathlib import Path
|
| 10 |
+
from typing import List, Dict, Any
|
| 11 |
+
import json
|
| 12 |
+
|
| 13 |
+
# Imports pour pyannote
|
| 14 |
+
try:
|
| 15 |
+
from pyannote.audio import Pipeline
|
| 16 |
+
HAS_PYANNOTE = True
|
| 17 |
+
except ImportError:
|
| 18 |
+
HAS_PYANNOTE = False
|
| 19 |
+
|
| 20 |
+
# Imports pour Whisper et Transformers
|
| 21 |
+
try:
|
| 22 |
+
import whisper
|
| 23 |
+
import torch
|
| 24 |
+
HAS_WHISPER = True
|
| 25 |
+
except ImportError:
|
| 26 |
+
HAS_WHISPER = False
|
| 27 |
+
|
| 28 |
+
try:
|
| 29 |
+
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
|
| 30 |
+
HAS_TRANSFORMERS = True
|
| 31 |
+
except ImportError:
|
| 32 |
+
HAS_TRANSFORMERS = False
|
| 33 |
+
|
| 34 |
+
# Corriger le problème PyTorch 2.6 avec weights_only
|
| 35 |
+
if hasattr(torch.serialization, 'add_safe_globals'):
|
| 36 |
+
try:
|
| 37 |
+
torch.serialization.add_safe_globals([torch.torch_version.TorchVersion])
|
| 38 |
+
except:
|
| 39 |
+
pass
|
| 40 |
+
|
| 41 |
+
import numpy as np
|
| 42 |
+
import librosa
|
| 43 |
+
import soundfile as sf
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def convert_audio_if_needed(audio_path: str) -> str:
|
| 47 |
+
"""
|
| 48 |
+
Convertit l'audio en WAV si nécessaire.
|
| 49 |
+
|
| 50 |
+
Returns:
|
| 51 |
+
Chemin vers le fichier audio (WAV si conversion nécessaire)
|
| 52 |
+
"""
|
| 53 |
+
ext = Path(audio_path).suffix.lower()
|
| 54 |
+
supported_formats = {'.wav', '.flac', '.ogg'}
|
| 55 |
+
|
| 56 |
+
if ext in supported_formats:
|
| 57 |
+
return audio_path
|
| 58 |
+
|
| 59 |
+
if ext in {'.m4a', '.mp3', '.mp4', '.aac'}:
|
| 60 |
+
wav_path = str(Path(audio_path).with_suffix('.wav'))
|
| 61 |
+
if os.path.exists(wav_path):
|
| 62 |
+
return wav_path
|
| 63 |
+
|
| 64 |
+
try:
|
| 65 |
+
y, sr = librosa.load(audio_path, sr=16000, mono=True)
|
| 66 |
+
sf.write(wav_path, y, sr)
|
| 67 |
+
return wav_path
|
| 68 |
+
except Exception as e:
|
| 69 |
+
return audio_path
|
| 70 |
+
|
| 71 |
+
return audio_path
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
def run_diarization(audio_path: str, hf_token: str, model_name: str = "pyannote/speaker-diarization-community-1") -> List[Dict[str, Any]]:
|
| 75 |
+
"""Exécute la diarisation avec pyannote."""
|
| 76 |
+
if not HAS_PYANNOTE:
|
| 77 |
+
raise ImportError("pyannote.audio n'est pas installé")
|
| 78 |
+
|
| 79 |
+
# Convertir l'audio en WAV si nécessaire
|
| 80 |
+
audio_path_converted = convert_audio_if_needed(audio_path)
|
| 81 |
+
|
| 82 |
+
# Configurer le token
|
| 83 |
+
if hf_token:
|
| 84 |
+
try:
|
| 85 |
+
from huggingface_hub import login
|
| 86 |
+
login(token=hf_token, add_to_git_credential=False)
|
| 87 |
+
except Exception:
|
| 88 |
+
pass
|
| 89 |
+
|
| 90 |
+
try:
|
| 91 |
+
pipeline = Pipeline.from_pretrained(model_name, token=hf_token)
|
| 92 |
+
except Exception as e:
|
| 93 |
+
if "plda" in str(e).lower() or "unexpected keyword" in str(e).lower():
|
| 94 |
+
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token=hf_token)
|
| 95 |
+
else:
|
| 96 |
+
raise
|
| 97 |
+
|
| 98 |
+
if torch.cuda.is_available():
|
| 99 |
+
pipeline = pipeline.to(torch.device("cuda"))
|
| 100 |
+
|
| 101 |
+
diarization = pipeline(audio_path_converted)
|
| 102 |
+
|
| 103 |
+
# Convertir en segments
|
| 104 |
+
segments = []
|
| 105 |
+
speakers = sorted(diarization.labels())
|
| 106 |
+
speaker_mapping = {speaker: f"SPEAKER_{idx:02d}" for idx, speaker in enumerate(speakers)}
|
| 107 |
+
|
| 108 |
+
for segment, track, speaker in diarization.itertracks(yield_label=True):
|
| 109 |
+
normalized_speaker = speaker_mapping.get(speaker, speaker)
|
| 110 |
+
segments.append({
|
| 111 |
+
"speaker": normalized_speaker,
|
| 112 |
+
"start": segment.start,
|
| 113 |
+
"end": segment.end
|
| 114 |
+
})
|
| 115 |
+
|
| 116 |
+
segments.sort(key=lambda x: x["start"])
|
| 117 |
+
return segments
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
def run_transcription(audio_path: str, device: str = None, hf_token: str = None) -> List[Dict[str, Any]]:
|
| 121 |
+
"""Exécute la transcription avec le modèle Whisper Large V3 French."""
|
| 122 |
+
if not HAS_WHISPER:
|
| 123 |
+
raise ImportError("whisper n'est pas installé")
|
| 124 |
+
|
| 125 |
+
if device is None:
|
| 126 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
| 127 |
+
|
| 128 |
+
model_id = "bofenghuang/whisper-large-v3-french"
|
| 129 |
+
|
| 130 |
+
# Utiliser Transformers pour charger le modèle
|
| 131 |
+
try:
|
| 132 |
+
if HAS_TRANSFORMERS:
|
| 133 |
+
processor = AutoProcessor.from_pretrained(model_id, token=hf_token)
|
| 134 |
+
model = AutoModelForSpeechSeq2Seq.from_pretrained(
|
| 135 |
+
model_id,
|
| 136 |
+
torch_dtype=torch.float16 if device == "cuda" else torch.float32,
|
| 137 |
+
low_cpu_mem_usage=True,
|
| 138 |
+
token=hf_token
|
| 139 |
+
)
|
| 140 |
+
model.to(device)
|
| 141 |
+
model.eval()
|
| 142 |
+
|
| 143 |
+
# Charger l'audio
|
| 144 |
+
audio_path_converted = convert_audio_if_needed(audio_path)
|
| 145 |
+
waveform, sample_rate = librosa.load(audio_path_converted, sr=16000, mono=True)
|
| 146 |
+
|
| 147 |
+
# Préparer les inputs
|
| 148 |
+
inputs = processor(
|
| 149 |
+
waveform,
|
| 150 |
+
sampling_rate=sample_rate,
|
| 151 |
+
return_tensors="pt"
|
| 152 |
+
)
|
| 153 |
+
inputs = {k: v.to(device) for k, v in inputs.items()}
|
| 154 |
+
|
| 155 |
+
# Transcription
|
| 156 |
+
with torch.no_grad():
|
| 157 |
+
generated_ids = model.generate(
|
| 158 |
+
inputs["input_features"],
|
| 159 |
+
language="fr",
|
| 160 |
+
task="transcribe",
|
| 161 |
+
return_timestamps=True
|
| 162 |
+
)
|
| 163 |
+
|
| 164 |
+
# Décoder avec timestamps
|
| 165 |
+
result = processor.batch_decode(
|
| 166 |
+
generated_ids,
|
| 167 |
+
skip_special_tokens=False,
|
| 168 |
+
output_word_timestamps=True
|
| 169 |
+
)[0]
|
| 170 |
+
|
| 171 |
+
# Extraire les segments avec timestamps depuis les tokens
|
| 172 |
+
tokens = generated_ids[0].cpu().numpy()
|
| 173 |
+
segments = []
|
| 174 |
+
current_segment = {"start": None, "end": None, "text": []}
|
| 175 |
+
|
| 176 |
+
# Parser les tokens pour extraire les timestamps
|
| 177 |
+
for token_id in tokens:
|
| 178 |
+
token_text = processor.tokenizer.decode([token_id], skip_special_tokens=False)
|
| 179 |
+
|
| 180 |
+
# Chercher les tokens de timestamp <|X.XX|>
|
| 181 |
+
if "<|" in token_text and "|>" in token_text:
|
| 182 |
+
try:
|
| 183 |
+
start_idx = token_text.find("<|") + 2
|
| 184 |
+
end_idx = token_text.find("|>")
|
| 185 |
+
if start_idx < end_idx:
|
| 186 |
+
timestamp_str = token_text[start_idx:end_idx]
|
| 187 |
+
timestamp = float(timestamp_str)
|
| 188 |
+
|
| 189 |
+
if current_segment["start"] is None:
|
| 190 |
+
current_segment["start"] = timestamp
|
| 191 |
+
else:
|
| 192 |
+
current_segment["end"] = timestamp
|
| 193 |
+
text = " ".join(current_segment["text"]).strip()
|
| 194 |
+
if text:
|
| 195 |
+
segments.append({
|
| 196 |
+
"start": current_segment["start"],
|
| 197 |
+
"end": current_segment["end"],
|
| 198 |
+
"text": text
|
| 199 |
+
})
|
| 200 |
+
current_segment = {"start": timestamp, "end": None, "text": []}
|
| 201 |
+
except (ValueError, IndexError):
|
| 202 |
+
pass
|
| 203 |
+
else:
|
| 204 |
+
if token_text.strip() and not any(x in token_text for x in ["<|", "|>", "<|startof", "<|endof", "<|notimestamps"]):
|
| 205 |
+
current_segment["text"].append(token_text)
|
| 206 |
+
|
| 207 |
+
# Ajouter le dernier segment
|
| 208 |
+
if current_segment["text"]:
|
| 209 |
+
text = " ".join(current_segment["text"]).strip()
|
| 210 |
+
if text:
|
| 211 |
+
duration = len(waveform) / sample_rate
|
| 212 |
+
segments.append({
|
| 213 |
+
"start": current_segment["start"] if current_segment["start"] is not None else 0.0,
|
| 214 |
+
"end": current_segment["end"] if current_segment["end"] is not None else duration,
|
| 215 |
+
"text": text
|
| 216 |
+
})
|
| 217 |
+
|
| 218 |
+
# Si on n'a pas réussi à extraire les timestamps, utiliser une approche de fallback
|
| 219 |
+
if not segments or all(seg.get("start") is None for seg in segments):
|
| 220 |
+
# Décoder le texte complet
|
| 221 |
+
result_text = processor.decode(generated_ids[0], skip_special_tokens=True)
|
| 222 |
+
|
| 223 |
+
# Diviser en phrases
|
| 224 |
+
sentences = []
|
| 225 |
+
for sent in result_text.split('. '):
|
| 226 |
+
if sent.strip():
|
| 227 |
+
sentences.append(sent.strip() + ('.' if not sent.strip().endswith('.') else ''))
|
| 228 |
+
|
| 229 |
+
if not sentences:
|
| 230 |
+
sentences = [result_text.strip()]
|
| 231 |
+
|
| 232 |
+
# Créer des segments temporels basés sur la durée
|
| 233 |
+
duration = len(waveform) / sample_rate
|
| 234 |
+
segments = []
|
| 235 |
+
time_per_sentence = duration / len(sentences)
|
| 236 |
+
|
| 237 |
+
for i, sentence in enumerate(sentences):
|
| 238 |
+
start_time = i * time_per_sentence
|
| 239 |
+
end_time = min((i + 1) * time_per_sentence, duration)
|
| 240 |
+
segments.append({
|
| 241 |
+
"start": start_time,
|
| 242 |
+
"end": end_time,
|
| 243 |
+
"text": sentence
|
| 244 |
+
})
|
| 245 |
+
|
| 246 |
+
return segments
|
| 247 |
+
except Exception as e:
|
| 248 |
+
# Fallback sur Whisper natif
|
| 249 |
+
model = whisper.load_model("large-v3", device=device)
|
| 250 |
+
|
| 251 |
+
audio_path_converted = convert_audio_if_needed(audio_path)
|
| 252 |
+
result = model.transcribe(
|
| 253 |
+
audio_path_converted,
|
| 254 |
+
language="fr",
|
| 255 |
+
task="transcribe",
|
| 256 |
+
verbose=False
|
| 257 |
+
)
|
| 258 |
+
|
| 259 |
+
segments = []
|
| 260 |
+
for seg in result["segments"]:
|
| 261 |
+
segments.append({
|
| 262 |
+
"start": seg["start"],
|
| 263 |
+
"end": seg["end"],
|
| 264 |
+
"text": seg["text"].strip()
|
| 265 |
+
})
|
| 266 |
+
|
| 267 |
+
return segments
|
| 268 |
+
|
| 269 |
+
|
| 270 |
+
def combine_diarization_transcription(
|
| 271 |
+
diarization_segments: List[Dict[str, Any]],
|
| 272 |
+
transcription_segments: List[Dict[str, Any]]
|
| 273 |
+
) -> List[Dict[str, Any]]:
|
| 274 |
+
"""Combine diarisation et transcription."""
|
| 275 |
+
combined = []
|
| 276 |
+
|
| 277 |
+
# Créer une timeline de diarisation
|
| 278 |
+
diar_timeline = [
|
| 279 |
+
(seg["start"], seg["end"], seg["speaker"])
|
| 280 |
+
for seg in diarization_segments
|
| 281 |
+
]
|
| 282 |
+
diar_timeline.sort()
|
| 283 |
+
|
| 284 |
+
def get_speaker_for_segment(seg_start: float, seg_end: float) -> str:
|
| 285 |
+
"""Détermine le locuteur pour un segment."""
|
| 286 |
+
speaker_time = {}
|
| 287 |
+
|
| 288 |
+
for diar_start, diar_end, speaker in diar_timeline:
|
| 289 |
+
overlap_start = max(seg_start, diar_start)
|
| 290 |
+
overlap_end = min(seg_end, diar_end)
|
| 291 |
+
overlap_duration = max(0, overlap_end - overlap_start)
|
| 292 |
+
|
| 293 |
+
if overlap_duration > 0:
|
| 294 |
+
speaker_time[speaker] = speaker_time.get(speaker, 0) + overlap_duration
|
| 295 |
+
|
| 296 |
+
if speaker_time:
|
| 297 |
+
return max(speaker_time, key=speaker_time.get)
|
| 298 |
+
else:
|
| 299 |
+
# Trouver le locuteur le plus proche
|
| 300 |
+
center_time = (seg_start + seg_end) / 2.0
|
| 301 |
+
min_dist = float('inf')
|
| 302 |
+
closest_speaker = "SPEAKER_00"
|
| 303 |
+
for diar_start, diar_end, speaker in diar_timeline:
|
| 304 |
+
if center_time < diar_start:
|
| 305 |
+
dist = diar_start - center_time
|
| 306 |
+
elif center_time >= diar_end:
|
| 307 |
+
dist = center_time - diar_end
|
| 308 |
+
else:
|
| 309 |
+
return speaker
|
| 310 |
+
if dist < min_dist:
|
| 311 |
+
min_dist = dist
|
| 312 |
+
closest_speaker = speaker
|
| 313 |
+
return closest_speaker
|
| 314 |
+
|
| 315 |
+
# Combiner les segments
|
| 316 |
+
for trans_seg in transcription_segments:
|
| 317 |
+
speaker = get_speaker_for_segment(trans_seg["start"], trans_seg["end"])
|
| 318 |
+
combined.append({
|
| 319 |
+
"speaker": speaker,
|
| 320 |
+
"start": trans_seg["start"],
|
| 321 |
+
"end": trans_seg["end"],
|
| 322 |
+
"text": trans_seg["text"]
|
| 323 |
+
})
|
| 324 |
+
|
| 325 |
+
return combined
|
| 326 |
+
|
| 327 |
+
|
| 328 |
+
def format_output(combined_segments: List[Dict[str, Any]]) -> str:
|
| 329 |
+
"""Formate la sortie en texte lisible: "Speaker A : blabla"."""
|
| 330 |
+
output_lines = []
|
| 331 |
+
|
| 332 |
+
current_speaker = None
|
| 333 |
+
current_texts = []
|
| 334 |
+
|
| 335 |
+
for seg in combined_segments:
|
| 336 |
+
speaker = seg["speaker"]
|
| 337 |
+
text = seg["text"]
|
| 338 |
+
|
| 339 |
+
if speaker != current_speaker:
|
| 340 |
+
# Écrire le groupe précédent
|
| 341 |
+
if current_speaker and current_texts:
|
| 342 |
+
speaker_num = int(current_speaker.replace("SPEAKER_", ""))
|
| 343 |
+
speaker_name = f"Speaker {chr(65 + speaker_num)}"
|
| 344 |
+
output_lines.append(f"{speaker_name} : {' '.join(current_texts)}")
|
| 345 |
+
|
| 346 |
+
# Nouveau locuteur
|
| 347 |
+
current_speaker = speaker
|
| 348 |
+
current_texts = [text]
|
| 349 |
+
else:
|
| 350 |
+
# Même locuteur, ajouter le texte
|
| 351 |
+
current_texts.append(text)
|
| 352 |
+
|
| 353 |
+
# Écrire le dernier groupe
|
| 354 |
+
if current_speaker and current_texts:
|
| 355 |
+
speaker_num = int(current_speaker.replace("SPEAKER_", ""))
|
| 356 |
+
speaker_name = f"Speaker {chr(65 + speaker_num)}"
|
| 357 |
+
output_lines.append(f"{speaker_name} : {' '.join(current_texts)}")
|
| 358 |
+
|
| 359 |
+
return "\n\n".join(output_lines)
|
| 360 |
+
|
requirements.txt
CHANGED
|
@@ -2,7 +2,15 @@ gradio>=4.0.0
|
|
| 2 |
pyannote.audio>=3.0.0
|
| 3 |
pyannote.core>=5.0.0
|
| 4 |
torch>=2.0.0
|
|
|
|
| 5 |
librosa>=0.10.0
|
| 6 |
soundfile>=0.12.0
|
| 7 |
huggingface-hub>=0.20.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
|
|
|
|
| 2 |
pyannote.audio>=3.0.0
|
| 3 |
pyannote.core>=5.0.0
|
| 4 |
torch>=2.0.0
|
| 5 |
+
torchaudio>=2.0.0
|
| 6 |
librosa>=0.10.0
|
| 7 |
soundfile>=0.12.0
|
| 8 |
huggingface-hub>=0.20.0
|
| 9 |
+
transformers>=4.30.0
|
| 10 |
+
openai-whisper>=20231117
|
| 11 |
+
accelerate>=0.20.0
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
|
| 16 |
|