Datasets:
license: cc-by-nc-4.0
task_categories:
- text-to-speech
language:
- en
tags:
- text-to-speech
size_categories:
- n<1K
Model Card for VoXtream2 Emilia dev dataset
This repository contains a dev dataset for the VoXtream2 TTS model, referred to as Emilia speaking-rate in the paper.
Audio prompts are derived from the Emilia dataset. We selected 62 speakers (balanced by gender), half of whom include filler words in the prompt (uniformly distributed across genders). Each speaker provides three audio prompts with slow (2.0 ± 0.5 SPS), normal (4.0 ± 0.6 SPS), and fast (5.6 ± 0.5 SPS) speaking rates. The prompt length distribution is similar to SEED test-en, with a mean of 4.5 seconds. Texts were generated using ChatGPT (2 per audio prompt), prompted with conversational-style sentences from the selected speakers and instructed to produce stylistically similar sentences containing 30-40 syllables (approximately 9-10 seconds of speech at a moderate speaking rate).
- Note: SPS - syllables per second.
Description
- prompt-audio - Audio prompts from Emilia-EN dataset (24kHz, flac format).
- metadata.csv - Test protocol for Emilia speaking-rate test, listing audio prompts with speaking rate characteristics and corresponding sentences to generate.
Usage
To download the dataset, use the following code:
from huggingface_hub import snapshot_download
local_dir = snapshot_download('voxtream2/emilia-dev', repo_type='dataset')