synth-data-tts / README.md
S24131's picture
Card: Hindi + English only
bc48b2b verified
|
Raw History Blame Contribute Delete
1.74 kB
---
license: other
task_categories:
- text-to-speech
language:
- hi
- en
pretty_name: Synthetic TTS data (Maya2-native)
configs:
- config_name: hindi
data_files:
- split: satvik
path: hindi/data/satvik/*.parquet
- split: charvi
path: hindi/data/charvi/*.parquet
default: true
- config_name: english
data_files:
- split: satvik
path: english/data/satvik/*.parquet
- split: charvi
path: english/data/charvi/*.parquet
---
# Synthetic TTS data
Synthetic speech generated with Maya2-native for distilling into a small TTS (Kokoro-82M). Two voices, Satvik
(male) and Charvi (female), per language. 24 kHz mono FLAC (16-bit) stored inside Parquet.
| Language | Voice | Clips | Hours | Shared texts | Shards |
|---|---|---|---|---|---|
| Hindi | Satvik | 140,002 | 242.2 | 33,003 | 36 |
| Hindi | Charvi | 140,002 | 261.4 | 98,003 | 38 |
| English | Satvik | 117,648 | 198.0 | 35,295 | 30 |
| English | Charvi | 117,647 | 210.5 | 35,295 | 32 |
One config per language, one split per voice. "Shared" clips use texts read by both voices; the rest are unique to
the voice. Texts are balanced by category and length bucket (`bucket`); `extra_numbers` marks added number-heavy
sentences.
## Layout
```
<language>/data/<voice>/train-XXXXX-of-NNNNN.parquet
<language>/text/ full sentence pool, per-voice selections (+ Kokoro train/val split for Hindi)
```
## Fields
`audio`, `id`, `voice`, `text` (original), `text_norm` (Maya normaliser output: numbers expanded, nukta stripped),
`category`, `source`, `bucket`, `shared`, `extra_numbers`, `duration` (s), `seed` (generation seed).
## Usage
```python
from datasets import load_dataset
ds = load_dataset("S24131/synth-data-tts", "hindi", split="satvik")
```