---
language:
- sna
license: mit
library_name: nemo
pipeline_tag: automatic-speech-recognition
tags:
- automatic-speech-recognition
- audio
- speech
- shona
- parakeet
- rnnt
- tdt
base_model:
- nvidia/parakeet-tdt-0.6b-v3
datasets:
- manassehzw/sna-dataset-annotated
metrics:
- wer
model-index:
- name: Shona Parakeet-TDT ASR
  results:
  - task:
      type: automatic-speech-recognition
      name: Automatic Speech Recognition
    dataset:
      name: manassehzw/sna-dataset-annotated
      type: manassehzw/sna-dataset-annotated
    metrics:
    - name: Validation Word Error Rate
      type: wer
      value: "N/A"
---

# manassehzw/sna-parakeet-tdt-0.6b-v3

Shona Parakeet-TDT ASR is a Shona (`sna`) automatic speech recognition model fine-tuned from
[`nvidia/parakeet-tdt-0.6b-v3`](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3) using the annotated dataset
[`manassehzw/sna-dataset-annotated`](https://huggingface.co/datasets/manassehzw/sna-dataset-annotated).

## Overview

This is a first-pass Shona fine-tune of NVIDIA Parakeet-TDT with especially promising
qualitative behavior on Shona-English code-switched speech. While the aggregate WER is not
yet state of the art for this dataset, the model produced strong real-world transcriptions
and preserved useful English recognition ability after Shona adaptation.

## Model Details

- **Curated by:** [Manasseh Changachirere](https://www.manasseh.dev/)
- **Base model:** [`nvidia/parakeet-tdt-0.6b-v3`](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3)
- **Training dataset:** [`manassehzw/sna-dataset-annotated`](https://huggingface.co/datasets/manassehzw/sna-dataset-annotated)
- **Run name:** `sna-parakeet-v1`
- **Uploaded artifact:** `sna-parakeet-v1.nemo`
- **Best validation WER during training:** N/A
- **Final validation WER:** N/A
- **Final test WER:** N/A

## Training Summary

- **Training examples:** 13715
- **Validation examples:** 762
- **Test examples:** 762
- **Valid rows after filtering:** 15239
- **Learning rate:** N/A
- **Microbatch size:** N/A
- **Gradient accumulation steps:** N/A
- **Effective batch size:** N/A
- **Max epochs:** N/A
- **Train max duration:** N/A
- **Validation max duration:** N/A

## Qualitative Notes

- Shona-only transcription is usable, with mostly soft orthographic substitutions.
- Shona-English code-switching remained surprisingly strong after fine-tuning.
- Real-world perceived quality appears stronger than the headline WER alone suggests.

## Limitations

- This is a first-pass fine-tune on a relatively small Shona corpus.
- The reported metrics come from the project’s held-out split, not from an external benchmark.
- Orthographic drift and near-phonetic substitutions are still common.
- Performance may differ on noisy audio, unseen accents, domain-shifted speech, and longer recordings.

## Files

This repository contains:

- the exported NeMo model artifact `sna-parakeet-v1.nemo`
- `training/summary.json`
- `training/preprocess_summary.json`
- `eval/validation_results.json`
- `eval/test_results.json`
- optional `eval/last_inference.json` qualitative sample output

## Example Usage

```python
import nemo.collections.asr as nemo_asr

repo_id = "manassehzw/sna-parakeet-tdt-0.6b-v3"
model = nemo_asr.models.ASRModel.from_pretrained(repo_id)
texts = model.transcribe(["sample.wav"])
print(texts[0].text if hasattr(texts[0], "text") else texts[0])
```