Xiaomi-CocktailASR-1 Technical Report

Humans have the remarkable ability to engage in spontaneous informal conversations and selectively attend to individual speakers while filtering out competing speech from nearby conversations. This "cocktail party" scenario still presents severe challenges to speech recognition systems. The CHiME-9 MCoRec task provides a testbed where systems must recognize groups of speakers and transcribe each of their conversations from audio-visual input. In this work, we analyze a diverse set of systems, representing different design directions for addressing the cocktail-party scenario, where the best system achieves up to 57% relative error reduction. We identify three main strategies: (1) explicit or implicit audio-visual target speech separation, (2) improved audio-visual speech recognition for each target speaker, and (3) the use of large language models to group speakers into conversations and enhance conversational consistency. Our analysis shows that these directions address complementary failure modes of the cocktail-party problem, and that high speech overlap alone does not explain performance differences, challenging the common assumption that overlap is the primary source of difficulty in cocktail-party recognition.

Multi-talker speech recognition is often addressed by combining automatic speech recognition (ASR) and speaker diarization in a pipeline system. Recently, LLM-based approaches have shown promise by jointly modeling semantic and speaker information, but they typically require large-scale multi-talker corpora that are costly to annotate. In this paper, we investigate how to efficiently train an LLM-based system with limited real-recorded data while maintaining high accuracy in speaker attribution. We propose several strategies: (1) a dual-encoder architecture to extract semantic and speaker features, (2) a feature interleaving format to merge these features as the inputs to the LLM, (3) a length-aware speaker ID loss to enhance diarization capability, and (4) an adaptive threshold strategy for ASR loss computation to mitigate hallucinations caused by speech overlaps. These strategies balance training between ASR and diarization tasks. Our system outperforms open-source baseline approaches, achieving relative improvements of 18% on the AliMeeting corpus and 24% on the Aishell4 corpus.

Large Language Models (LLMs) are effective decoders for Serialized Output Training (SOT) in two-talker automatic speech recognition (ASR), but their performance degrades substantially in three-talker mixtures. A key limitation is that conventional systems provide acoustic evidence only through an initial projected prefix, requiring the decoder to preserve fine-grained talker information throughout autoregressive generation. We first revisit CTC-derived static prefix conditioning using discrete token, hybrid token-acoustic, and continuous acoustic prompts. Although continuous acoustic cues are more reliable than discrete CTC hypotheses, the improvements on Libri3Mix remain limited, showing that richer prefix content alone does not resolve the conditioning bottleneck. We therefore propose persistent grounding in serialized acoustic memory, which enables the decoder to retrieve talker-structured acoustic evidence throughout the utterance. Specifically, talker-disentangled and onset-ordered acoustic representations are retained as external memory and accessed during decoding through gated residual cross-attention. We further introduce joint low-rank refinement of the acoustic retrieval pathway and selected LLM self-attention projections using LoRA. Experiments on Libri2Mix and Libri3Mix under clean and noisy conditions show consistent improvements over conventional LLM-SOT and naive stacked cross-attention, with particularly large gains in three-talker mixtures. These results demonstrate the importance of persistent access to structured, talker-aware acoustic evidence for LLM-based multi-talker ASR.