Audio

​

ASR Diarization

The ASR and diarization pipelines are modularly implemented, with diarization built on ASR outputs. Pyannote is recommended for state-of-the-art diarization. Speculative decoding is added to speed up inference, using a smaller model to suggest generations validated by the larger model. Note that this requires matching decoder architectures and a batch size of 1. For Whisper, a distilled version is suggested as the assistant model.

indexify-extractor download tensorlake/asrdiarization

​

Whisper

This extractor converts extracts transcriptions from audio. The entire text and chunks with timestamps are represented as metadata of the content.

​

Speaker Diarization

This extractor indentifies the speaker for each sentence in the transcription generated by Whisper.