Qwen3-ASR Atypical English — Projector Adaptation + SpecAugment
This repository contains the validation-selected projector checkpoint from the promptless Qwen3-ASR atypical-English architectural ablation.
Experiment
- Base model:
Qwen/Qwen3-ASR-1.7B - Dataset:
cdli/ugandan_english_nonstandard_speech_v1.0 - Condition:
projector - Prompt: none
- SpecAugment: enabled, train-only
- LR:
5e-05 - 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 | 3,148,800 |
| Trainable % | 0.1545% |
Validation checkpoint selection
Selection was performed on the validation split only using mean normalized per-utterance WER/CER capped at 1.0.
| Metric | Value |
|---|---|
| Selected checkpoint | checkpoint-750 |
| Mean capped WER | 33.61% |
| Mean capped CER | 24.14% |
| Corpus WER | 32.68% |
| Corpus CER | 22.40% |
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 | 24.17% |
| Mean capped utterance CER | 15.62% |
| Normalized corpus WER | 25.46% |
| Normalized corpus CER | 16.21% |
| Relative primary WER reduction vs base | 27.74% |
Results by severity
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|---|---|---|---|---|---|
| Mild (easily understood with minimal effort) | 334 | 3 | 20.91% | 12.27% | 24.30% | 13.77% |
| Moderate (requires effort to understand) | 340 | 3 | 24.31% | 15.00% | 23.34% | 14.45% |
| Severe (frequent breakdowns) | 339 | 3 | 27.23% | 19.55% | 28.61% | 20.63% |
Results by disorder
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|---|---|---|---|---|---|
| Articulation Disorders | 176 | 2 | 22.83% | 12.56% | 24.38% | 13.11% |
| Dysarthria | 253 | 2 | 25.05% | 15.60% | 24.78% | 15.48% |
| Stuttering (Disfluency Disorders) | 488 | 4 | 22.78% | 15.53% | 24.45% | 16.86% |
| Voice disorder | 96 | 1 | 31.35% | 21.79% | 32.28% | 22.35% |
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-en-atypical-promptless-specaug-results
No raw participant audio is redistributed.