---
language:
- en
task_categories:
- text-to-speech
tags:
- synthetic
- tts
- expressive-speech
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata.jsonl
---

# Maya1 expressive English TTS dataset (synthetic)

**Samples:** 37 | **Total audio:** 0.062 hours | **Mean duration:** 6.04 s | **Format:** 24000 Hz WAV

## Purpose and generation method
Synthetic speech for TTS research. Each record was produced in two stages: a small instruction model
(`Qwen/Qwen3-0.6B`) wrote an English utterance and inline-tag placement; the pipeline assigned controlled voice attributes
and rendered them as a voice description. Maya1 (Q8_0 GGUF, `audio-cpp/audio.cpp-gguf/Maya1-GGUF/maya1-q8_0.gguf`) then ran through the audio.cpp runtime
(v0.9.0, CUDA) to render the audio.
Voice design uses a controlled attribute vocabulary (realistic: age, gender, accent, pitch, timbre, pacing, emotion and intensity, domain, role, register; creative: character plus the same base attributes) rendered into a fixed descriptive string passed via `instruct=`. The sampled attributes are stored in `voice_attributes`.

## Scope
English only (default). 24 kHz WAV, sample rate and duration validated per file after generation.

## Inline tag convention
Tags are angle-bracket tokens placed inline in `text` at the point they should affect delivery.
Allowlist (20): <angry>, <appalled>, <chuckle>, <cry>, <curious>, <disappointed>, <excited>, <exhale>, <gasp>, <giggle>, <gulp>, <laugh>, <laugh_harder>, <mischievous>, <sarcastic>, <scream>, <sigh>, <sing>, <snort>, <whisper>. Many records contain no tag. `inline_tags` lists the tags actually found in `text`.

## Schema (`metadata.jsonl`)
`id`, `text`, `audio` (relative path), `duration_seconds`, `sample_rate`, `channels`, `language`, `voice_description`,
`voice_design_id`, `voice_attributes`, `topic`, `style`, `inline_tags`, `model_name`, `quantization`, `model_filename`, `runtime_version`,
`qwen_model`, `voice_description_source`, `generation_settings`, `seed`, `generation_wall_seconds`, `generation_status`, `created_at`.
`texts.jsonl` holds all accepted text records (some may have no audio yet); `generation_errors.jsonl` logs failed attempts.

## Loading
```python
from huggingface_hub import snapshot_download
from datasets import load_dataset, Audio
root = snapshot_download("Shadow0482/test-v1", repo_type="dataset")
ds = load_dataset("json", data_files=root + "/metadata.jsonl", split="train")
ds = ds.map(lambda r: {"audio": root + "/" + r["audio"]}).cast_column("audio", Audio(sampling_rate=24000))
print(ds[0]["text"], ds[0]["voice_description"])
```

## Known limitations
All text and audio are model-generated: pronunciation, prosody, accent and tag realisation can deviate from the written
description; invented content is not factual. Automatic validation is heuristic (duration, silence, format), not human review.
Near-duplicate filtering is lexical. Quality of a 0.6B-parameter text generator is limited.

## Licenses and notices
Maya1 is Apache-2.0 and its bundled SNAC decoder is MIT (as stated in the task brief; confirm on the upstream model cards).
See the model cards of `Qwen/Qwen3-0.6B` and the audio.cpp repository for their terms. **No license has been declared for this
dataset**; review upstream model/output terms before choosing one.

## Synthetic-data disclosure
This dataset consists entirely of synthetic speech and synthetic text. It does not contain recordings of real people.