DiarizationLM: Speaker Diarization Post-Processing with Large Language Models

What is DiarizationLM?

DiarizationLM is an approach for improving the output of a speaker diarization system by using a large language model as a post-processing step. Speaker diarization is the task of determining which parts of an audio recording were spoken by which speaker, typically assigning anonymous labels such as Speaker A and Speaker B rather than identifying anyone by name. DiarizationLM takes a preliminary diarization result and works to refine speaker-turn boundaries and labels, with the goal of producing a more coherent sequence of who spoke when.

How DiarizationLM fits into a multi-speaker audio pipeline

A typical multi-speaker pipeline may first separate or process the audio, transcribe the speech, and then apply diarization to associate speech segments with anonymous speakers. DiarizationLM works after an initial diarization pass. It treats the speaker-turn sequence similarly to how a language model treats a sequence of tokens, so the model can adjust speaker labels or turn structure based on learned patterns. This makes it useful when the original system produces fragmented turns or inconsistencies in speaker assignment.

Key concepts and scope

DiarizationLM is concerned with the task of labeling speaker turns. It is distinct from identifying who a person is, from transcribing the words they say, and from separating overlapping audio sources in the first place. The output uses anonymous speaker labels, and there is no documented evidence in the supplied material that it performs biometric identification or speaker recognition by name.

What the evidence does and does not establish

The available evidence defines DiarizationLM as a post-processing method for speaker diarization that relies on large language models. However, the supplied material does not include independent experimental results, speaker counts, accuracy scores, or performance measurements that could be used to compare DiarizationLM with other diarization systems. Any claims about specific performance improvements should be treated as undocumented unless supported by a further source.

Practical considerations for use

Before using DiarizationLM, it is important to know the limitations of the evidence available here. The approach may be useful for cleaning up initial speaker-turn predictions, but there are currently no documented measurements in this context to indicate how much improvement to expect. If you plan to integrate it into a system, prepare your own evaluation with a clear test set and confirm that speaker-turn labeling is the actual task you need rather than source separation, transcription, or identity recognition.