---
language:
- ta
task_categories:
- automatic-speech-recognition
tags:
- audio
- speech
- nemo
- manifest
dataset_info:
  splits:
  - name: train
  - name: test
  features:
  - name: audio_filepath
    dtype: string
  - name: duration
    dtype: float64
  - name: text
    dtype: string
  - name: target_lang
    dtype: string
---

# Audio Clips Dataset

NeMo-style ASR manifest dataset generated from `pod_mba`.

The dataset stores clipped WAV files under numbered `wavs/` subfolders and JSONL
manifests at `train_manifest.json` and `test_manifest.json`. Rows are
deterministically split into train/test with approximately 1% in
`test`.

## Columns

- `audio_filepath`: relative path to the clipped WAV file
- `duration`: clip duration in seconds
- `text`: transcript text
- `target_lang`: target language tag, set to `ta-IN`

WAV files are bucketed into numbered folders with up to 1000
files per folder, for example `wavs/000/...wav`, `wavs/001/...wav`, and so on.

## Usage

```bash
python your_nemo_training_script.py \
  --train_manifest train_manifest.json \
  --test_manifest test_manifest.json
```

Each WAV is mono audio resampled to 16000 Hz.