The SMIIP System for Multi-Speaker Automatic Speech Recognition in the 2nd MLC-SLM Challenge

Abstract

We present the SMIIP system for Task 1 of the 2nd Multi-lingual Conversational Speech Language Model (MLC-SLM) Challenge, which requires jointly performing speaker diarization and speech recognition on raw conversational recordings without oracle speaker labels or time boundaries. Our system couples a w2v-S2SND diarization front-end with DM-ASR, a diarization-aware multi-speaker ASR framework that reformu-lates transcription as a multi-turn dialogue generation process. The predicted speaker labels and utterance boundaries are discretized into dedicated speaker-index and time-index tokens, and each dialogue turn instructs a Qwen3-based 4B LLM to transcribe one speaker within one time segment, so that structural cues (who and when) are explicitly decoupled from lexical content (what). DM-ASR requires no speaker embedding extraction and conditions the LLM purely on discrete structural tokens. Our final system achieves a tcpMER of 24.50% and 19.97% on the development and evaluation sets, respectively, substantially outperforming the official baseline.