gabrielclark3330/expresso_ex04_ftext_audio

Emotion-Tagged Expresso Read ex04

A single-speaker, emotion-tagged repackaging of the official Expresso read-speech corpus for voice and expressive TTS fine-tuning. This repository contains all 2,903 transcribed WAV files for speaker ex04 (2.71 hours) as one training split.

Each transcript has exactly one style prefix followed by a space:

[confused] Why are you beating up my jukebox?

After that prefix, the official transcript is preserved exactly. Expresso markup such as *emphasis*, <laugh>, <breath>, and <singing> is neither removed nor normalized.

Styles

• confused: 380
• default: 621
• enunciated: 380
• happy: 380
• laughing: 380
• narration: 1
• sad: 380
• singing: 2
• whisper: 379

Columns

• audio: embedded original mono 48 kHz WAV bytes
• text: [emotion] followed by the untouched official transcript
• emotion: source Expresso read style
• source_file: original WAV filename
• duration_seconds: clip duration

The official benchmark splits are intentionally combined here so each repository is the complete corpus for one voice. The source material is the Expresso dataset, distributed under CC BY-NC 4.0 for non-commercial use.

Citation

@inproceedings{nguyen2023expresso,
  title={EXPRESSO: A Benchmark and Analysis of Discrete Expressive Speech Resynthesis},
  author={Nguyen, Tu Anh and Hsu, Wei-Ning and D'Avirro, Antony and Shi, Bowen and Gat, Itai and Fazel-Zarani, Maryam and Remez, Tal and Copet, Jade and Synnaeve, Gabriel and Hassid, Michael and Kreuk, Felix and Adi, Yossi and Dupoux, Emmanuel},
  booktitle={Interspeech},
  year={2023}
}