Stage 2 fine-tune on the CDLI Kenyan Swahili Nonstandard Speech dataset. Built on top of ElizabethMwangi/whisper-large-v3-swahili-afrivoice-cv (Stage 1).

Training Details

Dataset: cdli/kenyan_swahili_nonstandard_speech_v1.0
Stage 1 checkpoint: ElizabethMwangi/whisper-large-v3-swahili-afrivoice-cv
Encoder frozen: No
Decoder frozen: No
Projection frozen: No
SpecAugment: On
Learning rate: 5e-5 (polynomial decay)
Max steps: 1000
Best checkpoint: Step 1000
Output dir: sw_nss_tune_whisper_v3_large_afrivoice_cv_v1.2_redo

Results

Metric	Value
Avg WER	0.3011
Overall WER	0.2932
Avg CER	0.1289
Overall CER	0.1253
Overall WER (normalized)	0.311
Overall CER (normalized)	0.151
Avg WER (normalized)	0.269
Avg CER (normalized)	0.121

Per-Severity Results

Severity	WER (mean)	Count
Mild	0.203	92
Moderate	0.253	78
Severe	0.333	112

Usage

python

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="ElizabethMwangi/whisper-large-v3-swahili-nss-afrivoice-cv-v2"
)
result = asr("audio.wav")
print(result["text"])
Intended Use
This model is intended for transcription of Kenyan Swahili nonstandard speech, including dysarthric, stuttering, and otherwise atypical speech patterns.

Language token is set to Swahili (sw)

Citation
If you use this model, please cite the CDLI Kenyan Swahili Nonstandard Speech dataset and this repository.