Speaker Diarization

Speaker diarization divides audio into speaker turns and assigns each turn to a speaker, often alongside word-level transcription. Common approaches detect speech, create speaker embeddings, and cluster them; end-to-end neural models, streaming systems, and speech-language models instead learn speaker activity or joint transcription and attribution. Systems are tested on meeting and conversational benchmarks such as AMI, AliMeeting, and DIHARD, as well as multilingual, clinical, and long-form recordings. A central challenge is maintaining accurate identities through overlap, long recordings, and domain shifts while meeting latency constraints; annotation conventions and evaluation metrics can also change how errors are measured.