---
license: cc-by-4.0
language:
- lt
library_name: nemo
pipeline_tag: automatic-speech-recognition
base_model: nvidia/parakeet-tdt-0.6b-v3
datasets:
- Digisensus/lithuanian-dialect-speech-liepa-3-100h-punctuated
- Digisensus/lithuanian-phone-speech-liepa-3-429h-punctuated
- Digisensus/lithuanian-dialect-asr-eval
tags:
- speech
- lithuanian
- dialects
- parakeet
- tdt
---

# parakeet-tdt-0.6b-lt-study-s-d80-dial

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) + LIEPA-3 dialect speech (82.30 h), dialect clips in dialect spelling.
Output is lowercase spoken form without punctuation. This is a study model, not a product model.

## WER (%)

| Test | WER |
|---|---|
| Dialect test, dialect-spelling reference | 31.28 |
| Dialect test, standard-spelling reference | 36.29 |
| Dialect test, either spelling accepted | 27.60 |
| LIEPA-3 telephone test | 10.80 |
| FLEURS lt test | 23.24 |
| Common Voice 19 lt test | 25.77 |

## Use

```python
import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.from_pretrained("Digisensus/parakeet-tdt-0.6b-lt-study-s-d80-dial")
print(model.transcribe(["audio.wav"])[0].text)
```

## Provenance

| | |
|---|---|
| Run | `ltd26/S+D80.dial/s1/a3` |
| Recipe | `S+D80.dial` (`recipes/S+D80.dial.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 |