In this paper, we present OpenS2S, a fully open-source, transparent, end-to-end large-scale language model (LSLM) for empathic voice interaction. OpenS2S achieves low-latency speech generation using a streaming interleaved decoding architecture based on the empathic speech-to-text model BLSP-Emo. It integrates an automated data construction pipeline that synthesizes diverse, high-quality, empathic voice conversations at low cost, facilitating end-to-end learning. We leverage large-scale language models to generate empathic content, and introduce speaker and emotional variation using a controllable text-to-speech system, creating a scalable training corpus with rich paralinguistic diversity and minimal human supervision. We release the fully open-source OpenS2S model, including the dataset, model weights, pretraining, and fine-tuning code, to support the broader research community and accelerate innovation in empathic voice systems.
