---
language:
- en
task_categories:
- text-to-speech
tags:
- synthetic
- speech
- voice-cloning
- tts
size_categories:
- n<1K
---

# Synthetic single-speaker speech dataset (OmniVoice voice clone)

40 utterances, 0.07 hours of audio, 24000 Hz, mono, PCM 16-bit WAV.

## Voice-cloning disclosure
All audio is **synthetic speech** generated by a text-to-speech model that clones the voice of a reference
recording supplied by the dataset author. The speech is not a recording of a human reading these sentences.
Only use or redistribute it if the reference speaker has consented to this use and your use complies with the
upstream model licenses. Do not use it to impersonate people or to deceive.

## How it was made
1. **Text**: generated with `Qwen/Qwen2.5-0.5B-Instruct` (instruction prompts over varied topics, tones and forms; filtered for
   length, digits, symbols, boilerplate, a coarse unsafe-word list and near-duplicates). The text is model-written
   and may contain factual errors; it is intended as speech material, not as facts.
2. **Fine-grained control markup** (added by code after the text was written, not by the LLM):
   non-verbal tags (e.g. `[laughter]`, `[sigh]`) on about 35% of texts, balanced across tag types and
   placed at commas or the end of the sentence; pronunciation overrides on about 25% of texts
   (English: CMU phonemes for words with a single dictionary pronunciation; Chinese: pinyin with tone number).
   Observed in this dataset: 14 clips with a non-verbal tag ({'[confirmation-en]': 3, '[dissatisfaction-hnn]': 2, '[laughter]': 3, '[sigh]': 2, '[surprise-yo]': 1, '[question-en]': 1, '[question-oh]': 1, '[surprise-wa]': 1}), 10 clips with pronunciation overrides.
   `text_plain` is the same sentence without any markup.
3. **Speech**: OmniVoice 0.6B (`OmniVoice-0.6B (omnivoice-f16.gguf via audio.cpp v0.9.1)`) run through the audio.cpp runtime v0.9.1, cloning one reference voice
   (steps=20, guidance=2.0, speed=1.0, per-sample seed = 1234 + sample index).
4. **Validation**: each clip is decoded, checked for duration, silence, NaNs and a characters-per-second plausibility window,
   and stored together with its text. Duration is measured from the saved audio, not estimated.
   Whether the model actually honoured each tag or pronunciation override is NOT verified automatically.

## Layout
```
audio/000001.wav, ...
metadata.jsonl           one JSON object per clip
texts.jsonl              all generated texts (including texts not synthesized)
generation_errors.jsonl  failed or retried jobs
config.json              generation configuration (no secrets)
```

## Schema (`metadata.jsonl`)
| field | meaning |
|---|---|
| id | stable 6-digit sample id (matches the text id in `texts.jsonl`) |
| file_name, audio_path | relative audio path (`file_name` is what `datasets` AudioFolder reads) |
| text | the string that was synthesized (may contain tags / pronunciation overrides) |
| text_plain | the same sentence with all markup removed (use for ASR-style transcripts) |
| control_tags | non-verbal tags present in `text` |
| pron_overrides | list of pronunciation overrides applied (word/char and phonemes/pinyin) |
| duration_seconds, sample_rate | measured from the saved file |
| language, speaker_id, voice_description | language code, constant speaker id, free-text voice description |
| style, topic, form | prompt attributes used when the text was generated |
| model_name, text_model, seed, tts_steps | generation provenance |
| generation_status | `ok` |

## Loading
```python
from datasets import load_dataset
ds = load_dataset("Shadow0482/tts-v2", split="train")   # AudioFolder-style: metadata.jsonl + audio/
print(ds[0]["text"], ds[0]["audio"]["sampling_rate"])
```
(The loading snippet follows the standard `datasets` AudioFolder convention; it has not been executed by the generator.)

## Intended uses
Research and prototyping for TTS / voice-cloning pipelines, controllable-TTS experiments, speaker-adaptation studies.

## Limitations
* Single synthetic voice; prosody comes from a diffusion TTS model and may contain artifacts, mispronunciations or
  occasional truncation that automated checks cannot fully catch.
* Tags and pronunciation overrides are requests to the TTS model; the effect strength varies and is not measured.
* Small-LLM text: limited diversity, simple vocabulary, occasional awkward phrasing.
* The unsafe-word filter is coarse; review before use in sensitive settings.

## Licensing
No license is declared by this card because none has been verified for the combined artifact. Upstream components:
* OmniVoice weights via the audio.cpp GGUF package (`audio-cpp/audio.cpp-gguf`; the repo's table lists Apache-2.0 for this package) - verify before redistribution.
* Text model `Qwen/Qwen2.5-0.5B-Instruct` (Qwen2.5 Instruct small models are published under Apache-2.0 on their model cards) - verify.
* audio.cpp runtime: Apache-2.0.
The reference speaker's rights are separate and are the responsibility of the dataset author.

## Generation summary
Valid audio: 0.068 h (target 0.05 h, delta +0.018 h); clips: 40; mean duration 6.123 s; this session: 0:05:28 wall time, 0.7x realtime aggregate. Failed/pruned records are in generation_errors.jsonl.