README.md

---
language:
- sw
license: cc0-1.0
task_categories:
- text-to-speech
pretty_name: Swahili (swa_spk3) SNAC-Tokenized TTS Dataset for Orpheus Fine-Tuning
size_categories:
- 1K<n<10K
tags:
- tts
- speech
- swahili
- kiswahili
- snac
- orpheus
- audio-tokens
- low-resource
- african-languages
---

# Swahili (swa_spk3) SNAC-Tokenized Dataset for Orpheus-TTS Fine-Tuning

## Dataset Summary

A single-speaker Kiswahili subset, resampled and tokenized for fine-tuning [Orpheus-TTS](https://github.com/canopyai/Orpheus-TTS). It is derived from [`rlabz/swa_lug_tts`](https://huggingface.co/datasets/rlabz/swa_lug_tts) by:

1. Filtering the `train` and `validation` splits down to speaker **`swa_spk3`** only.
2. Resampling all audio from its original 22,050 Hz to **24,000 Hz**, the sample rate required by [SNAC](https://github.com/hubertsiuzdak/snac) (`snac_24khz`), the neural audio codec Orpheus is trained on.
3. Encoding each clip with SNAC into discrete audio codes and interleaving them with the text transcript into a single `input_ids` sequence, following the tokenization scheme used in the official [Orpheus fine-tuning notebook](https://github.com/canopyai/Orpheus-TTS).

The result is a training-ready dataset: no further audio processing is needed before feeding it into the Orpheus fine-tuning script.

- **Source speaker subset:** `swa_spk3` from `rlabz/swa_lug_tts`
- **Language:** Kiswahili (`sw`)
- **Audio codec / sample rate:** SNAC @ 24kHz
- **Intended use:** Fine-tuning Orpheus-TTS (Llama-3B-backbone) for a single Kiswahili voice
- **License:** [CC0 1.0 Public Domain](https://creativecommons.org/publicdomain/zero/1.0/) (inherited from the source dataset)

## Dataset Structure

### Data Splits

| Split | Utterances |
|---|---|
| train | 1,785 |
| validation | 198 |

### Data Fields

> ⚠️ The exact tokenization script used determines the precise field names — adjust this table if your pipeline's output differs.

| Field | Type | Description |
|---|---|---|
| `input_ids` | `list[int]` | Interleaved sequence of text token IDs (from the Llama tokenizer) and SNAC audio token IDs, following Orpheus's `<start_of_text> text <end_of_text> <start_of_speech> audio_codes <end_of_speech>` layout. Audio tokens are offset above the text vocabulary (IDs ≥ 128,000) so they occupy a disjoint range from text tokens. |
| `labels` | `list[int]` | Copy of `input_ids` used for next-token-prediction loss (standard causal LM fine-tuning target). |
| `attention_mask` | `list[int]` | Standard attention mask, all `1`s for non-padded sequences. |
| `speaker_id` | `string` | Always `swa_spk3` in this subset. |
| `language` | `string` | Always `swa`. |

### Data Instance

```python
{
  "input_ids": [128259, 264, 1495, ..., 128266, 7, 42, 91, ..., 128257],
  "labels": [128259, 264, 1495, ..., 128266, 7, 42, 91, ..., 128257],
  "attention_mask": [1, 1, 1, ...],
  "speaker_id": "swa_spk3",
  "language": "swa"
}
```

## Dataset Creation

### Source Data

Traces back to the [Luganda-Swahili Speech for Text-to-Speech Synthesis](https://www.kaggle.com/datasets/jocelyndumlao/luganda-swahili-speech-for-text-to-speechsynthesis/data) Kaggle dataset (CC0), processed into [`rlabz/swa_lug_tts`](https://huggingface.co/datasets/rlabz/swa_lug_tts) — see that dataset's card for details on corrupt-file filtering, speaker clustering, and the stratified train/validation split.

### Processing Steps

1. **Load** `rlabz/swa_lug_tts` and filter both `train` and `validation` splits to `speaker_id == "swa_spk3"` (1,785 train / 198 validation utterances).
2. **Resample** the `audio` column from 22,050 Hz to 24,000 Hz via `datasets.Audio(sampling_rate=24000)`, matching SNAC's expected input rate.
3. **Tokenize** each utterance with SNAC (`snac_24khz`) to produce hierarchical discrete audio codes, then interleave those codes with the text transcript's Llama tokenizer IDs into a single flat `input_ids` sequence, per the Orpheus fine-tuning data format.

### Why a single-speaker subset?

Orpheus fine-tuning for a specific voice is typically done on a single, consistent speaker rather than the full multi-speaker corpus, since mixing speakers in a single-voice fine-tune degrades voice consistency. `swa_spk3` was selected as the target voice for this fine-tune; the other 11 speakers in `rlabz/swa_lug_tts` remain available for separate single-speaker or multi-speaker experiments.

## Intended Use

This dataset is intended as direct input to the [Orpheus-TTS fine-tuning script](https://github.com/canopyai/Orpheus-TTS/tree/main/finetune) to produce a Kiswahili single-voice TTS model. It is not intended as a general-purpose ASR or multi-speaker TTS dataset — for that, use the source [`rlabz/swa_lug_tts`](https://huggingface.co/datasets/rlabz/swa_lug_tts) dataset instead.

## Licensing Information

Released under [CC0 1.0 Universal (Public Domain Dedication)](https://creativecommons.org/publicdomain/zero/1.0/), matching the license of the original Kaggle source and the parent `rlabz/swa_lug_tts` dataset.

## Citation

```bibtex
@misc{lugswa_tts_kaggle,
  title  = {Luganda-Swahili Speech for Text-to-Speech Synthesis},
  author = {Dumlao, Jocelyn},
  year   = {2024},
  url    = {https://www.kaggle.com/datasets/jocelyndumlao/luganda-swahili-speech-for-text-to-speechsynthesis/data}
}
```

Orpheus-TTS:

```bibtex
@misc{orpheus_tts,
  title  = {Orpheus-TTS: Towards Human-Sounding Speech},
  author = {Canopy Labs},
  year   = {2025},
  url    = {https://github.com/canopyai/Orpheus-TTS}
}
```

## Acknowledgements

Speaker filtering, resampling, and SNAC tokenization were performed as part of Orpheus fine-tuning data preparation under `rlabz`.
---