Maya TTS (maya-tts)

A high-fidelity expressive synthetic speech dataset for Maya, built using Qwen2.5-1.5B-Instruct for conversational dialogues and Fish Audio S1-mini for zero-shot voice cloning.

Dataset Overview

• Total Clips: 1518
• Total Duration: 1.451 hours (87.07 minutes)
• Audio Format: WAV (embedded directly into Parquet shards via Hugging Face Audio feature)
• Sampling Rate: 44100 Hz
• Language: English (en)

Features

Column	Type	Description
id	string	Unique zero-padded utterance identifier
audio	Audio	Embedded audio sample and decoding dict
text	string	Transcribed dialogue spoken by the cloned voice
emotion	string	Primary target emotion label (e.g. happy, anxious, excited)
tag	string	Specific delivery tone/special acoustic style (e.g. whispering, shouting)
language	string	Language code (en)
duration	float	Clip length in seconds

Reference Speaker Information

• Target Voice: Cloned from single reference prompt (ref_voice.wav).
• Reference Voice Transcript: "The fog had settled deep over the harbor long before midnight, turning the distant stone towers into jagged gray silhouettes against the sea. Out on the water, every sound carried double its daytime weight, the dull slap of cold black tide against the wooden pilings, the scrape of rusted iron chains, and the slow whistle of the gale driving through the cliffside pines. There is a particular stillness"

Distribution & Statistics

Top Emotions

Emotion	Count
satisfied	45
worried	39
scared	39
painful	38
sad	38
disdainful	38
surprised	38
delighted	37
sincere	37
angry	37

Delivery Styles & Tone Tags

Tag	Count
sobbing	41
laughing	39
sighing	38
chuckling	35
in a hurry tone	35
soft tone	34
screaming	34
whispering	32
crying loudly	29
shouting	27

Usage

from datasets import load_dataset

# Load directly as an iterable audio stream
dataset = load_dataset("Shadow0482/maya-tts", split="train", streaming=True)

for sample in dataset.take(2):
    print(sample["id"], sample["text"])
    audio_array = sample["audio"]["array"]
    sample_rate = sample["audio"]["sampling_rate"]

Generation Pipeline

• Script Generation: Prompted Qwen/Qwen2.5-1.5B-Instruct across multiple topics and affective directions.
• Acoustic Synthesis: Zero-shot target voice cloning via fishaudio/openaudio-s1-mini.
• Validation: Automatic duration, volume, and transcript alignment filtering.