New open-source pipeline called Sommelier aims to solve the data scarcity problem holding back full-duplex conversational AI systems

Sommelier is a scalable, open-source data processing pipeline designed specifically for training full-duplex Speech Language Models (SLMs) capable of real-time, natural two-way conversations

The pipeline addresses critical challenges in processing natural dialogue, including overlapping speech and back-channeling, which commonly cause diarization errors and ASR hallucinations in existing systems

Current full-duplex model development is severely limited by the lack of high-quality, multi-speaker conversational training data, as most large-scale resources contain only single-speaker or limited-volume content