Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hi-Fi Multi-Speaker English TTS Dataset

Evelina Bakhturina, Vitaly Lavrukhin, Boris Ginsburg, Yang Zhang

TL;DR

The paper presents Hi-Fi TTS, a high-quality, public domain, multi-speaker English TTS dataset derived from LibriVox and Gutenberg texts. It establishes strict audio quality criteria (bandwidth ≥13 kHz, SNR ≥32 dB) and uses zero-WER verification to ensure text–audio alignment, enabling more accurate and expressive TTS models. The dataset comprises 10 speakers with 17–58 hours each at 44.1 kHz, split into clean and other subsets, and includes dev/test splits with normalized references. The authors release the dataset and the accompanying NeMo-based processing pipeline to facilitate reproducible, high-fidelity multi-speaker TTS research.

Abstract

This paper introduces a new multi-speaker English dataset for training text-to-speech models. The dataset is based on LibriVox audiobooks and Project Gutenberg texts, both in the public domain. The new dataset contains about 292 hours of speech from 10 speakers with at least 17 hours per speaker sampled at 44.1 kHz. To select speech samples with high quality, we considered audio recordings with a signal bandwidth of at least 13 kHz and a signal-to-noise ratio (SNR) of at least 32 dB. The dataset is publicly released at http://www.openslr.org/109/ .

Hi-Fi Multi-Speaker English TTS Dataset

TL;DR

The paper presents Hi-Fi TTS, a high-quality, public domain, multi-speaker English TTS dataset derived from LibriVox and Gutenberg texts. It establishes strict audio quality criteria (bandwidth ≥13 kHz, SNR ≥32 dB) and uses zero-WER verification to ensure text–audio alignment, enabling more accurate and expressive TTS models. The dataset comprises 10 speakers with 17–58 hours each at 44.1 kHz, split into clean and other subsets, and includes dev/test splits with normalized references. The authors release the dataset and the accompanying NeMo-based processing pipeline to facilitate reproducible, high-fidelity multi-speaker TTS research.

Abstract

This paper introduces a new multi-speaker English dataset for training text-to-speech models. The dataset is based on LibriVox audiobooks and Project Gutenberg texts, both in the public domain. The new dataset contains about 292 hours of speech from 10 speakers with at least 17 hours per speaker sampled at 44.1 kHz. To select speech samples with high quality, we considered audio recordings with a signal bandwidth of at least 13 kHz and a signal-to-noise ratio (SNR) of at least 32 dB. The dataset is publicly released at http://www.openslr.org/109/ .

Paper Structure

This paper contains 11 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. The dataset construction
  3. Speaker selections
  4. Downloading metadata and audio samples
  5. Audio quality analysis
  6. Text pre-processing
  7. Audio segmentation and text-audio alignment
  8. Post-processing and sample selection
  9. Dataset splits and statistics
  10. Conclusions
  11. Acknowledgements

Figures (3)

  • Figure 1: Spectrograms of LibriVox recordings: (a) good audio quality (wideband signal, low noise); (b) not acceptable audio quality (narrowband signal); (c) not acceptable audio quality (low SNR); (d) not acceptable audio quality (both narrowband and noisy).
  • Figure 2: Energy-based VAD applied to a LibriVox recording.
  • Figure 3: Violin plots of the audio lengths per speaker: a) clean set, b) other set.