Project · 2026
Kikuyu Speech Recognition
Fine-tuned Whisper large-v3 on approximately 465.9 hours of Kikuyu speech. Audited transcription quality: 34.09% WER on the original test set and 25.01% on a retrospectively screened subset.
Overview
Fine-tuned Whisper large-v3 on approximately 465.9 hours of Kikuyu speech. Audited transcription quality: 34.09% WER on the original test set and 25.01% on a retrospectively screened subset.
Problem
Kikuyu speech recognition has limited transcribed training data, and inconsistent reference annotations can distort measured accuracy.
Approach
Fine-tuned Whisper large-v3 using approximately 465.9 hours of Kikuyu speech and speaker-disjoint data splits. Separately compared audio augmentation methods with Whisper-small and audited the evaluation references.
Results
The original held-out test returned 34.09% word error rate. A retrospective quality-screened subset returned 25.01% WER; this subset is not an untouched independent benchmark. The tested simple augmentations did not consistently improve the Whisper-small baseline.