Qwen3-ASR Atypical Luganda — Full Adaptation + SpecAugment
This repository contains the validation-selected full 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:
full
- 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 | 2,038,052,480 |
| Trainable % | 100.0000% |
Validation checkpoint selection
Selection was performed on the validation split only using
normalized corpus WER/CER.
| Metric | Value |
|---|
| Selected checkpoint | checkpoint-1250 |
| Mean capped WER | 65.47% |
| Mean capped CER | 31.81% |
| Corpus WER | 75.01% |
| Corpus CER | 34.06% |
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 | 58.79% |
| Mean capped utterance CER | 23.33% |
| Normalized corpus WER | 67.27% |
| Normalized corpus CER | 28.79% |
| Relative primary WER reduction vs base | 23.33% |
Results by severity
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|
| Mild (easily understood with minimal effort) | 366 | 3 | 53.25% | 18.06% | 60.98% | 23.73% |
| Moderate (requires effort to understand) | 347 | 3 | 58.79% | 21.82% | 63.06% | 22.64% |
| Severe (frequent breakdowns) | 315 | 3 | 65.21% | 31.13% | 79.52% | 43.09% |
Results by disorder
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|
| Articulation Disorders | 209 | 2 | 59.23% | 18.58% | 64.84% | 23.91% |
| Dysarthria | 276 | 2 | 53.33% | 19.43% | 58.02% | 22.71% |
| Stuttering (Disfluency Disorders) | 438 | 4 | 58.76% | 24.33% | 64.52% | 25.87% |
| Voice disorder | 105 | 1 | 72.35% | 38.92% | 93.06% | 53.70% |
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.
