OtoSpeech: 141 Hours of Processed Full-Duplex Conversational Audio

Otoearth released this 141-hour dataset of processed, two-speaker full-duplex conversational English speech in February 2026. It is a curated subset of the otoSpeech-full-duplex-280h collection, refined through human quality reviews and noise reduction techniques.

Use Cases

• Training audio-to-audio models for real-time interaction using the two-speaker full-duplex audio format
• Evaluating speech enhancement models by comparing the processed audio against raw conversational baselines
• Developing speaker diarization systems capable of handling natural overlaps in the 141 hours of dialogue

Strengths

• 141 hours of audio
• Human-reviewed for conversation quality
• Pre-processed with noise reduction and speech enhancement

Limitations

• Geographic bias toward US English speakers
• Potential selection bias due to human filtering for 'high quality'
• Unknown file formats and metadata structure

Provenance

Source

otoearth

Collection Method

Curated from a larger 280h dataset via human review and automated audio processing.

Freshness

Last updated February 2026.

Geography

United States