Modèle ASR Whisper,... pour la transcription audio en Bambara
Project description
whisper-bambara-sts
Modèle ASR (Automatic Speech Recognition) Whisper Small fine-tuné pour la transcription audio en bambara.
📊 Performance
| Métrique | Valeur |
|---|---|
| WER (Word Error Rate) | 37.38% |
| CER (Character Error Rate) | 16.45% |
| Dataset | 50h de bambara (63,813 échantillons) |
| Modèle de base | OpenAI Whisper Small (244M paramètres) |
| Taille du modèle | ~480 MB (FP16) |
🚀 Installation
pip install whisper-bambara-sts
Options d'installation
# Avec dépendances de développement
pip install whisper-bambara-sts[dev]
# Avec Gradio pour l'interface web
pip install whisper-bambara-sts[gradio]
# Toutes les dépendances
pip install whisper-bambara-sts[all]
💡 Utilisation
from whisper_bambara_sts import BambaraASR
# Initialisation (charge le modèle en mémoire)
asr = BambaraASR() # mode="best" par défaut (beam search)
# Transcription simple
text = asr.transcribe("audio.wav")
print(text)
# → "Sira e tɔgɔ"
# Mode rapide (greedy decoding)
asr_fast = BambaraASR(mode="fast")
text = asr_fast.transcribe("audio.wav")
# Avec surcharge des paramètres
text = asr.transcribe("audio.wav", num_beams=10)
# Avec numpy array
import numpy as np
audio = np.random.randn(16000 * 5).astype(np.float32) # 5 secondes
text = asr.transcribe(audio, sampling_rate=16000)
# Batch processing
files = ["audio1.wav", "audio2.wav", "audio3.wav"]
texts = asr.transcribe_batch(files)
CLI
# Transcription d'un fichier
bambara-sts audio.wav
# Mode rapide
bambara-sts audio.wav --mode fast
# Batch processing
bambara-sts --batch ./audios/
# Avec sortie fichier
bambara-sts audio.wav --output result.json
# Version
bambara-sts --version
Configuration
|Mode |num_beams |no_repeat_ngram_size| Vitesse | Qualité | |-----------------------------------------------------------------------------------------| |best (défaut) |5 |3 | Standard |Optimale (WER 37.38%)| |fast | 1 |3 | ~3x plus rapide |Légèrement inférieure|
Paramètres par défaut (verrouillés)
{
"max_new_tokens": 225,
"num_beams": 5, # beam search
"do_sample": False,
"no_repeat_ngram_size": 3, # évite répétitions
"early_stopping": True,
}
Ces paramètres ont été validés sur le test set complet (7,978 échantillons).
-
Formats audio supportés : WAV (recommandé), MP3, FLAC, OGG, M4A.
-
Spécifications recommandées : Sample rate : 16 kHz (auto-resampling si différent) Canaux : Mono (auto-conversion si stéréo) Bit depth : 16-bit ou 32-bit float
-
Cas d'usage:
- Applications mobiles : Transcription hors-ligne
- Santé : Transcription de consultations médicales
- 📰 Médias : Sous-titrage automatique
- 🎓 Éducation : Transcription de cours
- 🏛️ Administration : Numérisation d'archives audio
- 🌍 ONG : Collecte de données terrain au Mali
Developpement
# Cloner le repo
git clone https://github.com/kalilou/whisper-bambara-sts.git
cd whisper-bambara-sts
# Installer en mode développement
pip install -e ".[dev]"
# Lancer les tests
pytest tests/
# Formater le code
black src/whisper_bambara_sts/
ruff check src/whisper_bambara_sts/
📄 Licence:
MIT License - voir le fichier LICENSE pour plus de détails.
Remerciements:
- OpenAI pour le modèle Whisper
- La communauté bambara pour les données
- Hugging Face pour les outils Transformers
Contact
GitHub Issues : https://github.com/kalilou/whisper-bambara-sts/issues
MIT License Copyright (c) 2026 Kalilou Sangaré Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file whisper_bambara_sts-1.0.3.tar.gz.
File metadata
- Download URL: whisper_bambara_sts-1.0.3.tar.gz
- Upload date:
- Size: 13.4 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.14.2
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
f7d21993bb1dea25253dd49ec672f06effa9e983f6d54dcc1da05cce9766ca1b
|
|
| MD5 |
3d836cc7c65663eb18da9f406346cb74
|
|
| BLAKE2b-256 |
63fbe95c08e62c42a20b6c3fa0f5477e5f3ef3db1c0068993472bce26b9f2c87
|
File details
Details for the file whisper_bambara_sts-1.0.3-py3-none-any.whl.
File metadata
- Download URL: whisper_bambara_sts-1.0.3-py3-none-any.whl
- Upload date:
- Size: 12.5 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/6.2.0 CPython/3.14.2
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
c2b96bb1ff5684ca02c4682c9705024daf59bc08563a9c8647bdc09b159dbc57
|
|
| MD5 |
6d5509d35c38c15e5a5fec2990a9bd8d
|
|
| BLAKE2b-256 |
d589b348e509bc62e332944f997efa87416bcdb5a34a905287834039251d5ee1
|