whisper-large-v3_finetuned_rwandan_kinyarwanda_nonstandard_speech_v1.0

This is a fine-tuned version of cdli/whisper-large-v3_finetuned_kinyarwanda_standard_speech_v1.0 for Kinyarwanda non-standard speech. It is part of CDLI's effort to make speech technology work for people whose speech is underserved by mainstream ASR systems.

All CDLI models and datasets can be found on CDLI's HuggingFace page.

Dataset

The model has been fine-tuned using cdli/rwandan_kinyarwanda_nonstandard_speech_v1.0, a dataset of speech samples of people living with impaired speech across a range of impairment severity levels and etiologies.

Training

The train split was used for training, and the dev split for selecting the best checkpoint. Note: This Whisper model was fine-tuned and is decoded using the Swahili (sw) language setting — out of all languages Whisper supports, the one most similar to Kinyarwanda.

All model parameters (encoder, decoder, and output projection) were fine-tuned, with SpecAugment enabled.

Evaluation

This model was evaluated on the test split of the dataset. Utterances longer than 30 seconds were excluded:

• Examples evaluated: 1169
• Speakers: 9

For decoding we ran Whisper with language=sw, task=transcribe, greedy search (num_beams=1, do_sample=False).

Results are compared against the unadapted base model cdli/whisper-large-v3_finetuned_kinyarwanda_standard_speech_v1.0, evaluated identically, to show the effect of fine-tuning on non-standard speech.

We report two complementary word error rate (WER) metrics, both computed on text normalized with Whisper's BasicTextNormalizer, after first applying a Kinyarwanda-specific apostrophe normalization to both references and predictions: apostrophe glyphs (’, ‘, ʼ, `, ´) are unified to ', then the elision apostrophe is removed and the tokens joined (e.g. n'abantu → nabantu):

• Standard (corpus-level) WER — the usual error rate, pooling all reference words and edit errors across the entire test set.
• Per-utterance averaged WER — WER computed separately for each utterance, each capped at 1.0, then averaged across utterances.

The per-utterance averaged WER bounds each utterance to [0, 1] and weights all utterances equally, so it reflects typical performance without a few catastrophic utterances dominating — but it is not a true error rate and isn't directly comparable to other published WER, hence we report the standard, corpus-level WER as well.

Results

Overall Results

• Adapted — this CDLI model, fine-tuned on non-standard speech.
• Unadapted — the base model it was fine-tuned from (here: cdli/whisper-large-v3_finetuned_kinyarwanda_standard_speech_v1.0).

Detailed Analysis

Aggregated results can hide important underlying patterns, so we also break the WER down by subset: per speaker, and — where speaker severity is available — per impairment severity group.

Results by impairment severity

All WER values below are the per-utterance averaged WER, first averaged per speaker and then averaged within each severity group. n_speakers and n_utterances are the number of speakers and test utterances in each group.

Results by speaker

Per-utterance averaged WER per speaker. n_utterances is the number of test utterances for that speaker.