Modelpublic

ElizabethMwangi/whisper-large-v3-luganda-nss-ml-v1

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

0likes6downloads

Model Card

Whisper Large V3 — Multilingual Luganda + Ugandan English NSS (ml-v1)

Stage 2 multilingual fine-tune on the CDLI Ugandan Nonstandard Speech datasets. Built on top of lgwaxaltunewhisperlarge_7 (Stage 1). Trained jointly on Luganda and Ugandan English NSS data. Evaluated on Luganda NSS test set.

Training Details

—Datasets: cdli/ugandanlugandanonstandardspeechv1.0 + cdli/ugandanenglishnonstandardspeechv1.0
—Stage 1 checkpoint: lgwaxaltunewhisperlarge7/bestmodel
—Encoder frozen: No
—Decoder frozen: No
—Projection frozen: No
—SpecAugment: On
—Data augmentation: Yes (Gaussian noise, gain perturbation, speed perturbation)
—Learning rate: 5e-5 (polynomial decay)
—Max steps: 1000
—Best checkpoint: Step 850

Results on Luganda NSS Test Set (checkpoint-850)

Severity	Baseline WER	Fine-tuned WER
Mild	0.89	0.50
Moderate	0.96	0.53
Severe	0.97	0.60

Etiology	Baseline WER	Fine-tuned WER
Acquired hearing impairment	0.98	0.67
Cerebral Palsy	0.95	0.43
Cerebral palsy - cerebral malaria	0.93	0.55
Developmental	0.97	0.54
Idiopathic	0.80	0.45
Structural (tongue tie)	0.96	0.65

—Overall WER (normalized): 0.745
—Overall CER (normalized): 0.290
—Avg WER (normalized): 0.529
—Avg CER (normalized): 0.182
—Meaning preservation: 85.1%