Qwen3-ASR Atypical Luganda — Decoder Adaptation + SpecAugment
This repository contains the validation-selected decoder 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:
decoder
- 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 | 1,720,574,976 |
| Trainable % | 84.4225% |
Validation checkpoint selection
Selection was performed on the validation split only using
normalized corpus WER/CER.
| Metric | Value |
|---|
| Selected checkpoint | checkpoint-500 |
| Mean capped WER | 65.25% |
| Mean capped CER | 33.30% |
| Corpus WER | 73.59% |
| Corpus CER | 35.96% |
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 | 59.81% |
| Mean capped utterance CER | 25.88% |
| Normalized corpus WER | 76.43% |
| Normalized corpus CER | 38.76% |
| Relative primary WER reduction vs base | 21.99% |
Results by severity
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|
| Mild (easily understood with minimal effort) | 366 | 3 | 54.06% | 20.86% | 67.69% | 30.44% |
| Moderate (requires effort to understand) | 347 | 3 | 58.11% | 23.08% | 72.96% | 33.45% |
| Severe (frequent breakdowns) | 315 | 3 | 68.36% | 34.79% | 90.72% | 56.10% |
Results by disorder
| Group | n | Speakers | Mean capped WER | Mean capped CER | Corpus WER | Corpus CER |
|---|
| Articulation Disorders | 209 | 2 | 59.95% | 21.87% | 76.37% | 35.15% |
| Dysarthria | 276 | 2 | 51.19% | 19.24% | 56.13% | 21.92% |
| Stuttering (Disfluency Disorders) | 438 | 4 | 61.40% | 27.75% | 77.83% | 40.45% |
| Voice disorder | 105 | 1 | 75.53% | 43.50% | 110.60% | 72.83% |
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.
