Modelpublic

KasuleTrevor/qwen3-asr-lg-atypical-promptless-specaug-encoder

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes14downloads

Model Card

Qwen3-ASR Atypical Luganda — Encoder Adaptation + SpecAugment

This repository contains the validation-selected encoder checkpoint from the promptless Qwen3-ASR atypical-Luganda architectural ablation.

Experiment

—Base model: KasuleTrevor/cdli-qwen3-asr-lg-typical-1p7b-base-finetune
—Dataset: cdli/ugandan_luganda_nonstandard_speech_v1.0
—Condition: encoder
—Prompt: none
—SpecAugment: enabled, train-only
—LR: 0.0001
—Scheduler: cosine
—Seed: 42
—Batch / grad accumulation: 4 / 4
—Effective batch: 16

SpecAugment

Setting	Value
Probability	0.5
Time masks	2
Time width	40
Frequency masks	2
Frequency width	12

Trainable parameters

Metric	Value
Total	2,038,052,480
Trainable	314,328,704
Trainable %	15.4230%

Validation checkpoint selection

Selection was performed on the validation split only using normalized corpus WER/CER.

Metric	Value
Selected checkpoint	`checkpoint-1500`
Mean capped WER	72.50%
Mean capped CER	39.35%
Corpus WER	100.80%
Corpus CER	60.06%

Final test results

Primary reporting uses normalized per-utterance WER/CER, capped at 1.0 per utterance and then averaged. Normalized corpus metrics are retained as secondary diagnostics.

Metric	Result
Mean capped utterance WER	67.71%
Mean capped utterance CER	31.38%
Normalized corpus WER	83.54%
Normalized corpus CER	44.98%
Relative primary WER reduction vs base	11.69%

Results by severity

Group	n	Speakers	Mean capped WER	Mean capped CER	Corpus WER	Corpus CER
Mild (easily understood with minimal effort)	366	3	60.66%	24.21%	76.46%	39.25%
Moderate (requires effort to understand)	347	3	65.30%	28.10%	76.28%	36.91%
Severe (frequent breakdowns)	315	3	78.57%	43.31%	100.24%	62.60%

Results by disorder

Group	n	Speakers	Mean capped WER	Mean capped CER	Corpus WER	Corpus CER
Articulation Disorders	209	2	68.41%	30.83%	91.40%	51.32%
Dysarthria	276	2	60.17%	25.05%	65.72%	29.26%
Stuttering (Disfluency Disorders)	438	4	68.45%	31.18%	84.72%	42.44%
Voice disorder	105	1	83.07%	49.92%	98.77%	64.39%

Reproducibility

The exact corrected trainer is archived in this repository under:

code/qwen3_asr_sft_fixed_ablation.py

The companion private results repository is:

KasuleTrevor/qwen3-asr-lg-atypical-promptless-specaug-results

No raw participant audio is redistributed.