Moonshine student m3g, 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 | 71.13% | 55.24% | 0 | 46 |
| beam 4 | 65.14% | 50.60% | 0 | 82 |
Exact transcripts 24.72%; substitutions 54.49%, deletions 8.94%, insertions 14.83% of reference words (uncapped WER 78.26%).
WER by duration
| bucket | recordings | WER | share of all errors |
|---|---|---|---|
| 1-2s | 670 | 53.45% | 0.29% |
| 2-3s | 4354 | 44.06% | 2.18% |
| 3-5s | 14700 | 47.91% | 10.40% |
| 5-8s | 14739 | 59.83% | 18.33% |
| 8-12s | 7023 | 72.07% | 15.07% |
| 12-20s | 3867 | 81.24% | 15.34% |
| 20-35s | 1657 | 88.48% | 16.62% |
| >35s | 919 | 91.30% | 21.76% |
WER by etiology
| etiology | recordings | WER | share of all errors |
|---|---|---|---|
| ALS | 10232 | 70.00% | 21.82% |
| Cerebral Palsy | 11472 | 73.46% | 20.09% |
| Down Syndrome | 7719 | 68.24% | 10.99% |
| Parkinson's Disease | 14543 | 72.95% | 42.28% |
| Stroke | 3963 | 62.11% | 4.83% |
Runaway transcripts (hypothesis over 1.5x the reference length): 2410, 12.33% of all errors; truncated transcripts: 904.
Training recipe
{
"init": "/runs/sapc_moonshine/inputs/m3g_cut_from_m5g",
"teacher": "/runs/sapc_moonshine/runs/m15s2/best",
"kd_weight": 0.5,
"peak_lr": 5e-05,
"epochs": 1,
"balanced": true,
"balance_max_repeats": 2,
"policy": {
"max_items": 768,
"padded_seconds": 4608,
"attention_cost": 60000,
"decoder_tokens": 98304,
"accumulation": 2
},
"seed": 20261030
}
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).