---
language:
- vi
license: cc-by-4.0
source_datasets:
- doof-ferb/LSVSC
task_categories:
- automatic-speech-recognition
- text-to-speech
pretty_name: a novel large-scale Vietnamese speech corpus (LSVSC)
tags:
- audio
- compressed
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
audio_compress:
- config: default
  split: train
  format: Opus
  settings: level=0.9
  columns:
  - audio
  source: doof-ferb/LSVSC
  status: done
  rows: 45458
  hours: 80.63
  size: 9.27 GB → 1.15 GB
---

# LSVSC (compressed audio)

> **This is a compressed version of [`doof-ferb/LSVSC`](https://huggingface.co/datasets/doof-ferb/LSVSC).**
> Only the audio column(s) were re-encoded (see the table below) to reduce download size;
> every other column, subset and split name is kept unchanged.

**License:** same as the original dataset (`cc-by-4.0`). Please refer to the [original dataset card](https://huggingface.co/datasets/doof-ferb/LSVSC) for the full description, citation and terms of use.

## Compressed splits

<!-- audio-compress:table:start -->
| config | split | format | settings | sample rate | rows | audio hours | size (orig → new) |
|---|---|---|---|---|---:|---:|---|
| `default` | `train` | Opus | level=0.9 | original | 45,458 | 80.63 | 9.27 GB → 1.15 GB |
<!-- audio-compress:table:end -->

## Usage

```python
from datasets import load_dataset

ds = load_dataset('giangndm/LSVSC-opus', 'default', split='train')
```

Lossy formats (Opus / MP3) are not bit-exact with the original audio. Encoded with [`soundfile`](https://github.com/bastibe/python-soundfile) (libsndfile).