Synthetic Speech Dataset with Seven Distribution Shifts

ShiftySpeech is a large-scale synthetic speech dataset containing over 3000 hours of audio. It was created by user ash56 and spans seven distinct distribution shifts including reading style, podcast, YouTube, three languages, and demographic variations. The dataset was last updated on HuggingFace in October 2025.

Use Cases

• Benchmark speech recognition model robustness by testing performance on the 'Podcast' versus 'Reading Style' speech segments.
• Analyze the impact of 'Demographics' features like age and accent on synthetic speech quality metrics.
• Train multi-lingual TTS systems using data from the 'Languages' shift covering three different languages.
• Study domain adaptation techniques using the 'YouTube' style speech data compared to other styles.

Strengths

• Over 3000 hours of synthetic speech audio
• Covers 7 distinct distribution shift categories
• Includes data synthesized from multiple TTS systems

Limitations

• Specific row count and file size are unknown
• Synthetic nature may not fully capture real-world speech characteristics
• Class balance and sample sizes per shift category are unspecified

Provenance

Source

ash56 on HuggingFace

Collection Method

Synthesized using various text-to-speech systems

Freshness

Last updated October 2025.