parakeet-tdt-0.6b-lt-study-s-x80

Research checkpoint from the study Lithuanian speech recognition: effects of dialect training and transcript spelling. Fine-tuned from nvidia/parakeet-tdt-0.6b-v3 on LIEPA-3 telephone speech (420.17 h) + other LIEPA-3 spontaneous speech (82.36 h) (control). Output is lowercase spoken form without punctuation. This is a study model, not a product model.

WER (%)

Test	WER
Dialect test, dialect-spelling reference	39.90
Dialect test, standard-spelling reference	33.86
Dialect test, either spelling accepted	32.18
LIEPA-3 telephone test	10.31
FLEURS lt test	24.95
Common Voice 19 lt test	34.07

Use

import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.from_pretrained("Digisensus/parakeet-tdt-0.6b-lt-study-s-x80")
print(model.transcribe(["audio.wav"])[0].text)

Provenance

Run	ltd26/S+X80/s1/a2
Recipe	S+X80 (recipes/S+X80.json in the code repo)
Splits	splits-v1
Training	10,000 updates, 192 clips per update, last checkpoint; full spec in config.json
Scorer	see registry/scores.csv in the code repo