Deepgram ships Batch Diarization V2 for pre-recorded speaker labeling
- Source
- x.com
- Date
Introducing Batch Diarization V2, a major upgrade to speaker labeling for pre-recorded audio. Understanding what was said is only part of the story. You also need to know who said it. Highlights: 🔹 Preferred 3.3X in human evaluation 🔹 Improved speaker attribution accuracy across real-world audio 🔹 Available today via the new diarize_model parameter Get started:
Context
Deepgram's Batch Diarization V2 updates the speaker labels attached to prerecorded speech. Those labels distinguish who said each part of a conversation, so mistakes can assign one person's words to another. The release article describes a new speaker model, improved segmentation and clustering, and a larger training set, which Deepgram says reduce wrongly merged or split speakers.
The advertised 3.3-fold preference is a human-comparison result: of 158 votes, 63.3% favored V2, 19.0% favored V1 and the rest had no preference. It is not a 3.3-fold accuracy improvement. Requests using diarize=true remain on V1; passing diarize_model=v2 explicitly selects the new version.
- embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
Checking sign-in…
Loading comments…