GigaSpeech: 10,000 Hours of Multi-Domain English Transcribed Audio

GigaSpeech is a multi-domain English speech recognition corpus containing 10,000 hours of high-quality labeled audio released by SpeechColab in 2021. The data is aggregated from audiobooks, podcasts, and YouTube, capturing a mix of read and spontaneous speaking styles across topics like arts, science, and sports.

Use Cases

• Training Automatic Speech Recognition (ASR) models using the transcribed audio segments
• Developing Text-to-Speech (TTS) synthesis using the high-quality audio-text pairs
• Domain-specific acoustic modeling for niche topics like science or sports

Strengths

• 10,000 hours of labeled training data
• Multi-domain coverage including arts, science, and sports
• Includes both read and spontaneous speaking styles

Limitations

• Monolingual English focus prevents use for cross-lingual tasks
• Potential for variable audio quality and background noise in YouTube and podcast sources

Provenance

Source

SpeechColab (Arxiv 2106.06909)

Collection Method

Scraped and transcribed from audiobooks, podcasts, and YouTube

Freshness

Last updated February 2026 according to platform metadata.

Geography

United States