Or4kool/mms-lid-256-fra-dyu
Spoken language identification for Côte d'Ivoire: classifies audio clips as french, dioula, or other. Fine-tuned from facebook/mms-lid-256.
Evaluation (held-out test split)
- fine-tuned clean test: accuracy 0.9979, macro-F1 0.9974
- fine-tuned telephony-simulated slice: macro-F1 0.9974
- zero-shot
facebook/mms-lid-256baseline (mapped to 3 classes): macro-F1 0.4459
| class | precision | recall | f1 | support |
|---|---|---|---|---|
| french | 0.9979 | 0.9986 | 0.9983 | 1429 |
| dioula | 1.0 | 0.9988 | 0.9994 | 840 |
| other | 0.9944 | 0.9944 | 0.9944 | 540 |
Usage
from transformers import pipeline
classifier = pipeline("audio-classification", model="Or4kool/mms-lid-256-fra-dyu")
print(classifier("clip.wav"))
Trained on public datasets only (Common Voice Dioula, GO AI Corp Dioula, FarmerlineML Dyula, African-accented French, FLEURS). Bambara was excluded from training; the other class approximates non-target languages with other African languages plus English.