Longform speaker diarization by prompting multimodal llm with chunk-wise in-context self-enrollment

Abstract

A method includes receiving a prompt including a textual diarization request and corresponding audio data characterizing a conversation between multiple speakers. The method also includes generating a sequence of audio encoding chunks based on the corresponding data. For each respective audio encoding chunk, the method includes using a trained large language model (LLM) generating corresponding diarization results based on the respective audio encoding chunk and the textual diarization request and generating a new audio cohort for the respective audio encoding chunk based on the corresponding diarization results. The corresponding diarization results include a speech recognition result that has one or more predicted terms. Each respective predicted term is associated with a corresponding speaker token representing a predicted identity of a respective speaker that spoke the respective predicted term. The trained LLM is conditioned on a prior audio cohort generated by the trained LLM for a prior audio encoding chunk.