README.md

metadata

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 Seamless test dataset

This repository contains a test dataset for the VoXtream2 TTS model, referred to as Seamless test in the paper.

Audio files are derived from the Seamless Interaction dataset. We build a test set from the naturalistic subset of the Seamless Interaction test split. We select 10-20-second sentences whose average speaking rates are uniformly distributed over 2-6 SPS, from 45% male and 55% female speakers. For each speaker, the 3-7-second acoustic prompt (4.7 seconds on average) is taken from another sentence with an average rate of 3-4 SPS. After removing sentences with a Whisper-large-v3 WER above 10%, we sample 322 unique sentences, totaling 1.1 hours.

• Note: SPS - syllables per second, WER - word error rate

Description

• prompt-audio - Audio prompts from Seamless interaction dataset (24kHz, flac format).
• audio - Audio files from Seamless interaction dataset (24kHz, flac format).
• metadata.csv - Test protocol for Seamless test, listing audio prompts 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/seamless', repo_type='dataset')