TC-BiMamba: Trans-Chunk bidirectionally within BiMamba for unified streaming and non-streaming ASR

Researchers introduced TC-BiMamba, a new architecture for unified streaming and non-streaming automatic speech recognition (ASR). It uses a novel Trans-Chunk mechanism for dynamic chunk size training, enabling a single model to handle both offline decoding and low-latency streaming. The method achieves a 1.3x training speedup, reduces training memory by 50%, and matches or outperforms prior models like U2++ and LC-BiMamba while using a smaller model size.