Expresso: 40 Hours of 48kHz Expressive Speech in 34 Styles

Expresso is a 40-hour expressive speech dataset featuring 48kHz high-fidelity audio from four speakers across 34 distinct styles. Developed by ylacombe and documented in Arxiv 2308.05725, the collection includes 11 hours of read speech and 30 hours of improvised dialogues. The data is balanced between two male and two female speakers and includes transcriptions for the read speech components.

Use Cases

• Expressive Text-to-Speech (TTS) synthesis using the 8 read styles
• Style and emotion classification across 26 improvised dialogue categories
• High-fidelity audio generation and vocoder training leveraging the 48kHz sampling rate

Strengths

• 48kHz high-fidelity audio sampling rate
• 40 total hours of speech data
• 34 distinct expressive styles including 26 improvised categories
• Includes both mono and stereo recording formats

Limitations

• Small speaker diversity with only 4 individuals represented
• Restricted to English language speech
• Non-commercial license limits industrial application

Provenance

Source

ylacombe (Arxiv 2308.05725)

Collection Method

Recorded expressive read speech and improvised dialogues from 4 speakers.

Freshness

Last updated in April 2024.

Geography

United States