LibriQuote: Speech Dataset of Fictional Character Utterances for Expressive Zero-Shot Speech Synthesis
This repository contains helper functions to process LibriQuote data, and benchmark expressive TTS systems using LibriQuote-test.
This repository contains:
- Helper python classes to process LibriQuote in the
processing/folder - Evaluation scripts to benchmark TTS systems on LibriQuote-test in the
evaluation/folder. - The LibriQuote dataset hosted on HuggingFace.
Figure 1. t-SNE projection of emotion vector representations computed with emotion2vec-plus-base. LibriQuote-test (a) quotations and (b) reference narration (non-quotation) utterances; (c) Subsample of LibriHeavy segments (N=5734).
If you use LibriQuote-test only for benchmarking, we provide target and reference samples (in 16KHz) directly in the HuggingFace repository. Check-out the evaluation/ folder to find evaluation scripts.
LibriQuote comes with segments derived from narration paragraphs and quotation from characters in fiction novels. It is derived from LibriVox recordings, and currently uses LibriLight audio files as backend audio files. Note that these audio files are encoded in 16KHz.
Please follow LibriLight instructions to download and prepare audio files.
We provide a bash script that will untar only necessary LibriQuote files, reducing the overall processing time and total disk space required
Find more information in the processing/ folder.
Find more information in the evaluation/ folder.
If you use LibriQuote or part of this code in your publications, you can cite this work with the following BibTex entry:
@misc{Michel2025LibriQuote,
title={LibriQuote: A Speech Dataset of Fictional Character Utterances for Expressive Zero-Shot Speech Synthesis},
author={Gaspard Michel and Elena V. Epure and Christophe Cerisara},
year={2025},
eprint={2509.04072},
archivePrefix={arXiv},
primaryClass={eess.AS},
url={https://arxiv.org/abs/2509.04072}
}