Qwen3-ASR Atypical English — Decoder Adaptation + SpecAugment
This repository contains the validation-selected decoder 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:
decoder
- 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 | 1,720,574,976 |
| Trainable % | 84.4225% |
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 | 33.38% |
| Mean capped CER | 23.95% |
| Corpus WER | 32.42% |
| Corpus CER | 22.25% |
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 | 22.46% |
| Mean capped utterance CER | 14.25% |
| Normalized corpus WER | 23.81% |
| Normalized corpus CER | 14.84% |
| Relative primary WER reduction vs base | 32.84% |
Results by severity
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|
| Mild (easily understood with minimal effort) | 334 | 3 | 19.61% | 11.74% | 22.81% | 13.25% |
| Moderate (requires effort to understand) | 340 | 3 | 23.15% | 14.06% | 22.46% | 13.65% |
| Severe (frequent breakdowns) | 339 | 3 | 24.57% | 16.92% | 26.11% | 17.76% |
Results by disorder
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|
| Articulation Disorders | 176 | 2 | 21.59% | 11.97% | 23.10% | 12.73% |
| Dysarthria | 253 | 2 | 23.70% | 14.64% | 23.76% | 14.63% |
| Stuttering (Disfluency Disorders) | 488 | 4 | 20.12% | 13.18% | 20.78% | 13.47% |
| Voice disorder | 96 | 1 | 32.67% | 22.86% | 35.01% | 24.42% |
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.
