---
language:
- km
license: cc-by-sa-4.0
task_categories:
- text-to-speech
- automatic-speech-recognition
pretty_name: Khmer TTS Processed (Fish Speech format)
size_categories:
- 10K<n<100K
---

# Khmer TTS Processed

Preprocessed, Fish Speech-ready Khmer speech data derived from
[DDD-Cambodia/khm-asr-cultural](https://huggingface.co/datasets/DDD-Cambodia/khm-asr-cultural)
(`train` split). Built for fine-tuning [Fish Speech](https://github.com/fishaudio/fish-speech)
(`fishaudio/openaudio-s1-mini`) for Khmer text-to-speech / voice cloning —
see [Panhapich/Tuna-TTS](https://huggingface.co/Panhapich/Tuna-TTS) for the
resulting model checkpoints.

## Processing pipeline

Raw audio + transcripts were pulled from the source dataset and run through:

1. Export/merge raw clips into a single manifest
2. Automated audio-quality (QC) grading
3. Denoising
4. VAD-based silence trimming
5. Loudness normalization
6. Khmer text normalization (numbers, currency, dates, percentages verbalized)
7. Per-speaker stratified train/validation split
8. Conversion to Fish Speech's speaker-folder format (paired `.wav` + `.lab`
   files, one folder per speaker)

Source code: [Pich09/voice-clone](https://github.com/Pich09/voice-clone),
`scripts/01_download_ddd.py` through `scripts/09_convert_to_fish_format.py`.

## Splits

| Split | Clips  | Speakers | Duration  |
|-------|--------|----------|-----------|
| train | 54,083 | 10       | 125.2 h   |
| valid | 1,123  | 6        | 2.6 h     |

## Repository layout

Each archive is a `.tar` of Fish Speech-format `.wav`/`.lab` pairs (or, for
the `ready_*` archives, already VQ-extracted protobuf training shards).

| File | Contents |
|---|---|
| `processed/khmer_base__khmer_base_v1.tar` | Full train + val set in one archive (raw audio, unsharded, ~22GB) |
| `processed/khmer_base__khmer_base_v1__shard{0..5}of6.tar` | Train set split into 6 raw-audio shards (~3.5GB each) |
| `processed/khmer_base__khmer_base_v1__val.tar` | Validation set (raw audio, unsharded, ~430MB) |
| `processed/khmer_base__khmer_base_v1__ready_shard{i}of6.tar` | VQ-extracted + packed protobuf training data for shard `i` (much smaller than raw audio — ready to train on directly, no VQ extraction needed) |
| `processed/khmer_base__khmer_base_v1__ready_val.tar` | VQ-extracted + packed protobuf validation data |
| `processed/khmer_base__khmer_base_v1__shards.json` | Shard-count manifest |
| `processed/khmer_base__khmer_base_v1__ready.json` | Tracks which shards have completed VQ preprocessing |

The shard split exists to fit Kaggle's 20GB `/kaggle/working` disk cap, which
can't hold the full ~22GB dataset at once — see
`kaggle/preprocess_khmer_vq_v2.ipynb` in the source repo for how shards are
VQ-preprocessed one at a time, and `kaggle/train_khmer_base.ipynb` /
`scripts/10_train_fish_khmer_base.sh` for how the data is consumed in
training.

## Audio quality

QC grading is automated (signal-based heuristics), not manually verified per
clip — treat grades as a filtering signal, not a guarantee of studio quality.

## License

The source dataset is licensed **CC BY-SA 4.0**. This derived/processed
version is distributed under the same license and its share-alike terms.

## Citation

Please cite the original dataset:

```
DDD-Cambodia/khm-asr-cultural, https://huggingface.co/datasets/DDD-Cambodia/khm-asr-cultural
```