Moonshine student m5g, SAPC2 compression ladder
Fine-tuned for the Speech Accessibility Project Challenge 2 (SAPC2): English speech from speakers with ALS, cerebral palsy, Down syndrome, Parkinson's disease and stroke.
Evaluation
The checkpoint was selected on a fixed 1,000-recording subset of the official Dev set; the scores below are on the full Dev set (47,929 recordings). Scores use the competition normalizer against the better of the two references; they are within about 0.1 points of the official SCTK scorer on the same predictions.
Full Dev (47,929 recordings)
Whole-recording decoding (200 s windows), quick scorer (competition normalizer, better of the two references).
| decoding | WER | CER | empty transcripts | capped (ran to the token limit) |
|---|---|---|---|---|
| greedy | 61.22% | 44.73% | 0 | 57 |
| beam 4 | 54.85% | 39.98% | 0 | 78 |
Exact transcripts 29.92%; substitutions 46.94%, deletions 7.39%, insertions 10.66% of reference words (uncapped WER 65.00%).
WER by duration
| bucket | recordings | WER | share of all errors |
|---|---|---|---|
| 1-2s | 670 | 49.96% | 0.31% |
| 2-3s | 4354 | 37.57% | 2.21% |
| 3-5s | 14700 | 40.36% | 10.37% |
| 5-8s | 14739 | 50.34% | 18.24% |
| 8-12s | 7023 | 61.99% | 15.37% |
| 12-20s | 3867 | 70.70% | 15.67% |
| 20-35s | 1657 | 76.50% | 16.34% |
| >35s | 919 | 80.07% | 21.50% |
WER by etiology
| etiology | recordings | WER | share of all errors |
|---|---|---|---|
| ALS | 10232 | 58.18% | 20.68% |
| Cerebral Palsy | 11472 | 65.74% | 20.90% |
| Down Syndrome | 7719 | 60.35% | 11.80% |
| Parkinson's Disease | 14543 | 62.47% | 42.07% |
| Stroke | 3963 | 52.05% | 4.56% |
Runaway transcripts (hypothesis over 1.5x the reference length): 1547, 7.15% of all errors; truncated transcripts: 693.
Training recipe
{
"init": "runs/from_modal/m5g/best",
"teacher": "runs/from_modal/m15s2/best",
"kd_weight": 0.5,
"peak_lr": 2e-05,
"epochs": 1,
"balanced": true,
"balance_max_repeats": 2,
"policy": {
"max_items": 192,
"padded_seconds": 1152,
"attention_cost": 12000,
"decoder_tokens": 24576,
"accumulation": 8
},
"seed": 20261026
}
Optimizer: blocked SOAP (Adam in a Shampoo eigenbasis), FP32 weights with BF16 autocast.
Code: the moonshine package in the project repository (training, distillation, RL and
evaluation), with resumable checkpoints.
Files
Hugging Face format (config.json, model.safetensors, tokenizer/processor files). Load with
transformers (MoonshineForConditionalGeneration or CohereAsrForConditionalGeneration).