README.md

metadata

license: cc-by-4.0
language:
  - en
task_categories:
  - text-to-speech
tags:
  - tts
  - expressive-speech
  - synthetic
  - maya1
  - voice-design
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

Maya1 Expressive TTS Dataset

A synthetic English speech dataset with inline expressive tags and a natural-language voice description for every clip.

Clips	3,850
Total audio	4.88 hours
Mean / median clip length	4.6 s / 4.1 s
Sample rate	24 kHz, mono WAV (embedded in parquet)
Language	English
Clips with at least one inline tag	25 (0.6%)

How it was made

1. Text: Qwen/Qwen3-0.6B wrote short spoken utterances (1-3 sentences) from a planned set of speaker labels, topics, tones and tag instructions. Texts were validated (allowed tags only, word-count limits, no duplicates).
2. Voice design: each clip has a structured voice plan (gender, age, accent, pitch, timbre, pacing, emotion, intensity, plus a domain/role or a creative character), rendered into the instruction string that Maya1 consumes.
3. Audio: maya1-q8_0 (GGUF) run through audio.cpp on Kaggle T4 GPUs. Clips were checked for sample rate, silence, duration and speech-rate anomalies, and failed ones were dropped.

Columns

column	description
id	clip id, e.g. maya1_000123
audio	WAV audio (24 kHz), decoded automatically by datasets
text	transcript, including inline tags such as <laugh>
duration_seconds	clip length in seconds
sample_rate	sample rate in Hz
language	English
voice_description	the instruction string given to Maya1
voice_attributes	JSON string of the structured voice plan
topic	short topic label
inline_tags	list of expressive tags in the text
n_words	word count of the transcript, tags excluded from nothing (plain split)
model	maya1-q8_0
seed	sampling seed used for this clip

Usage

from datasets import load_dataset
import json

ds = load_dataset("Shadow0482/test-v1", split="train")
ex = ds[0]
print(ex["text"])
print(ex["voice_description"])
print(json.loads(ex["voice_attributes"]))
audio = ex["audio"]            # {"array": ..., "sampling_rate": 24000}

Allowed inline tags: <<chuckle>>, <<curious>>, <<disappointed>>, <<exhale>>, <<gasp>>, <<giggle>>, <<sarcastic>>, <<sigh>>.

Distribution

voice kind	clips	share
realistic	2725	70.8%
creative	1125	29.2%

gender	clips	share
male	2027	52.6%
female	1823	47.4%

accent	clips	share
american	1521	39.5%
british	1020	26.5%
indian	549	14.3%
middle_eastern	392	10.2%
asian_american	368	9.6%

emotion	clips	share
neutral	2597	67.5%
dry	322	8.4%
energetic	262	6.8%
excited	259	6.7%
sad	252	6.5%
sarcastic	158	4.1%

inline tag	clips	share
	8	32.0%
	5	20.0%
	3	12.0%
	3	12.0%
	2	8.0%
	2	8.0%
	1	4.0%
	1	4.0%

Limitations

• All audio and text are synthetic. Prosody and tags come from a quantized (Q8_0) model and may be imperfect.
• Texts come from a 0.6B language model, so they can be repetitive or generic.
• Voice descriptions come from a fixed vocabulary, so coverage of real-world voices is limited.
• Check the license of Maya1 and Qwen before commercial use. The license field above is a placeholder: change it if needed.