Qwen3-ASR Atypical English — Full Adaptation + SpecAugment
This repository contains the validation-selected full 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:
full
- 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 | 2,038,052,480 |
| Trainable % | 100.0000% |
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-1250 |
| Mean capped WER | 32.74% |
| Mean capped CER | 23.65% |
| Corpus WER | 32.59% |
| Corpus CER | 22.65% |
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 | 22.79% |
| Mean capped utterance CER | 14.46% |
| Normalized corpus WER | 24.30% |
| Normalized corpus CER | 15.10% |
| Relative primary WER reduction vs base | 31.85% |
Results by severity
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|
| Mild (easily understood with minimal effort) | 334 | 3 | 19.95% | 11.70% | 23.30% | 13.10% |
| Moderate (requires effort to understand) | 340 | 3 | 22.89% | 13.89% | 22.38% | 13.67% |
| Severe (frequent breakdowns) | 339 | 3 | 25.49% | 17.75% | 27.10% | 18.72% |
Results by disorder
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|
| Articulation Disorders | 176 | 2 | 22.04% | 11.85% | 23.78% | 12.66% |
| Dysarthria | 253 | 2 | 24.22% | 14.99% | 24.06% | 14.83% |
| Stuttering (Disfluency Disorders) | 488 | 4 | 20.82% | 13.69% | 21.96% | 14.29% |
| Voice disorder | 96 | 1 | 30.42% | 21.81% | 33.28% | 23.68% |
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.
