full_moonshine_tiny_all_data_wup5_ep12_lr1e-4_new_transcripts
Full fine-tune of UsefulSensors/moonshine-tiny (27.1M parameters) for Macedonian speech-to-text, trained on the same all-data mix as full_moonshine_tiny_all_data_wup1_ep3_lr1e-4 (vezilka-asri plus seven curated sets -- videa_so_transkript_od_yt, alfa_audios, sitel_audios, doniraj (accepted donations), mozzila_common_voice, fleurs_mk and jargon), retrained after vezilka-asri's transcripts were corrected/retranscribed -- same data selection as that run, but the audio-text pairs themselves changed for the ~99% of the mix that vezilka-asri makes up.
Moonshine is a raw-waveform encoder-decoder: it takes input_values rather than
Whisper's fixed 30 s log-mel window, so its cost scales with actual audio length
instead of being flat per clip. That is what makes it viable on-device, and it is
why its real-time factor barely moves with utterance length where Whisper's does.
Evaluation
Held-out test set rachno_provereno_od_yt (44 clips, 0.32 h),
cuda, batch 16, 1 beam.
| metric | value |
|---|---|
| WER | 11.91 |
| CER | 4.85 |
| SER | 100.00 |
| Parameters | 27.1 M |
| Latency | 118 ms/utterance |
| Throughput | 223x real-time |
| RTF | 0.00448 |
SER is exact-match over whole utterances. These test clips average 62 words, so at this WER an exactly-correct utterance is vanishingly unlikely -- read WER/CER.
Best WER seen during training (on the same held-out set): 12.02
at step 535000 (epoch 5.65 of the 12 configured -- load_best_model_at_end
kept this checkpoint as the final save rather than a later, worse one).
Note: the peak learning rate actually used was 5e-5, not the lr1e-4 its
output-dir/repo name carries over from an earlier revision of the config it was
launched from -- the Training table below reflects the value actually used.
Compared with the doniraj/prvatv Whisper sweep
This model is far behind every Whisper LoRA/full-finetune adapter in that sweep (WER ~4-5% there vs 11.91% here) -- expected, since Moonshine-tiny is a 27M-parameter model built for on-device use, being compared against a fine-tuned 809M-parameter Whisper Large v3 Turbo.
Training
| setting | value |
|---|---|
| learning rate | 5e-05 |
| warmup ratio | 0.05 |
| epochs (configured / reached) | 12 / 5.65 |
| batch size | 32 x 1 accum |
| scheduler | cosine |
| weight decay | 0.01 |
| precision | bf16 |
| steps at best checkpoint | 535000 |
Usage
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
import librosa, torch
model_id = "h-gajdov/full_moonshine_tiny_all_data_wup5_ep12_lr1e-4_new_transcripts"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id).eval()
wav, _ = librosa.load("clip.wav", sr=16000, mono=True)
feats = processor.feature_extractor([wav], sampling_rate=16000,
return_tensors="pt", padding=True)
with torch.no_grad():
ids = model.generate(**feats, max_length=192)
print(processor.batch_decode(ids, skip_special_tokens=True)[0])
max_length is capped at 192 because the decoder has 194 positions
(max_position_embeddings). Moonshine has no forced language/task tokens, so do
not pass language= or task= to generate() -- it will raise.