The MOSS Systems for the MLC-SLM Challenge 2026

Abstract

This paper describes our systems for the Multilingual Conversational Speech Language Model (MLC-SLM) Challenge 2026. The challenge contains two tasks: multilingual conversational transcription with speaker attribution, and multiple-choice conversational speech understanding. For Task 1, we use MOSS Transcribe Diarize (MTD), an end-to-end speech language model that directly generates timestamped speaker-attributed transcripts from long conversational audio. MTD combines a Whisper-large-v3 audio encoder, an acoustic adaptor, and a Qwen3-8B decoder, and is adapted with multilingual ASR data, speaker-attributed conversational data, synthetic multi-speaker mixtures, and the official MLC-SLM training data. For Task 2, we adapt MOSS-Audio-8B-Thinking to answer multiple-choice questions directly from audio, us-ing challenge-style supervised and reinforcement-learning data. The submitted system obtains 14.93 tcpMER on the Task 1 evaluation set and 0.94 accuracy in both Task 2 evaluation phases.