Moonshine student m10s, 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 | 43.60% | 29.39% | 0 | 38 |
| beam 4 | 38.49% | 25.78% | 0 | 87 |
Exact transcripts 38.48%; substitutions 32.67%, deletions 5.59%, insertions 7.31% of reference words (uncapped WER 45.57%).
WER by duration
| bucket | recordings | WER | share of all errors |
|---|---|---|---|
| 1-2s | 670 | 40.32% | 0.34% |
| 2-3s | 4354 | 27.06% | 2.25% |
| 3-5s | 14700 | 27.93% | 10.09% |
| 5-8s | 14739 | 35.99% | 18.34% |
| 8-12s | 7023 | 45.95% | 15.96% |
| 12-20s | 3867 | 51.96% | 16.13% |
| 20-35s | 1657 | 53.43% | 15.92% |
| >35s | 919 | 55.39% | 20.97% |
WER by etiology
| etiology | recordings | WER | share of all errors |
|---|---|---|---|
| ALS | 10232 | 39.99% | 20.26% |
| Cerebral Palsy | 11472 | 52.37% | 23.47% |
| Down Syndrome | 7719 | 46.42% | 12.48% |
| Parkinson's Disease | 14543 | 41.68% | 39.23% |
| Stroke | 3963 | 37.93% | 4.56% |
Runaway transcripts (hypothesis over 1.5x the reference length): 926, 5.43% of all errors; truncated transcripts: 402.
Training recipe
{
"init": "/runs/sapc_moonshine/inputs/m10s_cut_from_m15s2",
"teacher": "/runs/sapc_moonshine/runs/m15s2/best",
"kd_weight": 0.5,
"peak_lr": 3e-05,
"epochs": 2,
"balanced": true,
"balance_max_repeats": 2,
"policy": {
"max_items": 768,
"padded_seconds": 4608,
"attention_cost": 60000,
"decoder_tokens": 98304,
"accumulation": 2
},
"seed": 20261022
}
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).