---
license: apache-2.0
language:
- en
task_categories:
- text-to-speech
tags:
- synthetic
- tts
- voice-cloning
- omnivoice
pretty_name: ic-tts
---

# ic-tts

Synthetic speech dataset: LLM-written text, spoken by **OmniVoice** (voice cloning from reference audio).

| | |
|---|---|
| Utterances | 395 |
| Total duration | 1.00 h |
| Mean / min / max length | 9.1s / 3.5s / 15.2s |
| Sample rate | 24000 Hz mono, FLAC |
| Language | English |
| Reference voices | 1 |

## Pipeline
1. Text: `Qwen/Qwen2.5-0.5B-Instruct` prompted with random topic/style/length combos, cleaned and de-duplicated -> `texts.json`.
2. Audio: OmniVoice (0.6B, GGUF f16) through [audio.cpp](https://github.com/0xShug0/audio.cpp) v0.9.1, zero-shot voice clone, 2x T4 in parallel.
3. Basic filtering: clips that are too short or whose duration is implausible for the text length are dropped.

## Files
- `audio/` - audio clips
- `metadata.jsonl` - one row per clip: `id`, `text`, `path`, `duration` (seconds), `sample_rate`, `ref_voice` (`file_name` duplicates `path` for the HF audio viewer)
- `texts.json` - all generated texts (including any not synthesized)

## Usage
```python
from datasets import load_dataset
ds = load_dataset("Shadow0482/ic-tts", split="train")
print(ds[0]["text"], ds[0]["duration"])
```

## Notes and limitations
- Fully synthetic. Text can contain LLM errors or odd phrasing; speech can contain TTS artifacts. No manual verification was done.
- The cloned voice(s) come from reference recordings. Only use voices you have the right/consent to clone, and check the licenses of the models used (Qwen, OmniVoice) before commercial use.