Synthetic Duplex Speech Conversations For Real-Time Dialogue Training

104,478 fully synthetic duplex conversations provide 2,133 hours of 16kHz audio for training real-time spoken dialogue models. The dataset was created by author mailong225 for the RelayS2S hybrid architecture, converting text dialogues to speech. It was last updated on March 25, 2026.

Use Cases

• Train end-to-end speech-to-speech models on 2,133 hours of synthetic duplex conversation audio for low-latency dialogue.
• Benchmark the quality of cascaded ASR-LLM pipelines against fully synthetic speech-to-speech training data.
• Develop hybrid architectures like RelayS2S that combine real-time S2S models with high-quality response pipelines using synthetic dialogues.

Strengths

• Contains 104,478 synthetic conversations, a substantial volume for model training.
• Offers 2,133 hours of 16kHz audio, providing extensive temporal coverage for speech tasks.
• Specifically designed for training real-time spoken dialogue systems, providing a focused application.

Limitations

• Data is fully synthetic, which may not capture the acoustic and conversational nuances of real human speech.
• Limited information on audio diversity, speaker characteristics, or underlying text dialogue sources.
• The dataset's utility for tasks beyond the specific RelayS2S architecture is untested.

Provenance

Source

Hugging Face user mailong225.

Collection Method

Constructed by converting text dialogues to speech to create synthetic duplex conversations.

Time Range

null

Freshness

Last updated March 25, 2026.

Geography

null