Expresso: 40-Hour Expressive Speech Dataset with Read and Improvised Styles

The Expresso dataset is a high-quality 48kHz expressive speech collection created by author somu9. It includes 40 hours of audio from 4 speakers, comprising 11 hours of expressively rendered read speech and 30 hours of improvised dialogues. The dataset was last updated on 2026-05-18.

Use Cases

• Training expressive text-to-speech models based on the 8 read speech styles.
• Developing conversational AI agents based on the 26 improvised dialogue styles.
• Researching prosody and emotion in speech based on the high-quality 48kHz audio recordings.
• Benchmarking speech synthesis models on a multi-speaker dataset with both scripted and spontaneous speech.

Strengths

• Contains 40 hours of total audio data, providing a substantial resource.
• Features high-quality 48kHz audio recordings.
• Includes data from 4 speakers (2 male, 2 female), offering speaker diversity.
• Provides transcriptions for the 11 hours of read speech.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2026-05-18 07:04:04; freshness should be verified.

Provenance

Source

huggingface user somu9

Freshness

Last updated 2026-05-18 07:04:04.