Speaker diarization
Core model parameters ~8.1M
Model files 30.95 MB
Transcribe a meeting, then analyze its speakers offline and replay each turn with linked text and a timeline
- Speaker labels
- Speaker timeline
- Linked replay
Recorded on the board
Screen recording of the board web demo, recorded on Qualcomm QCS8550. (Transcription and analysis waits are visibly sped up; replay is at normal speed. This is not an inference benchmark. Media: AMI Consortium, CC BY 4.0, credited in the video.)
Highlights
Transcribe, then analyze
Transcribe while the meeting video plays. After pausing or finishing the recording, start speaker analysis to finalize labels for that recording.
Explore by speaker
Colors and local speaker IDs distinguish turns. Filter by speaker; IDs apply only to the current recording and do not identify real people.
Linked text and timeline
Click transcript text or a timeline segment to locate and replay the corresponding turn.
Export the results
Export results as JSON, or analyze complete audio recordings through the C API and Python SDK.
Runs on Qualcomm chips
The board SDK handles transcription and speaker analysis without uploading meeting audio to a cloud service.
Review uncertain assignments
Unresolved and context-estimated assignments remain marked. Short recordings and weak word boundaries may require review. Speaker analysis is offline.
Pipeline
From input to output: where every step runs.
- 01Browser
Meeting input
Play a meeting video and capture audio for live transcription
- 02NPU
Live transcription
On-device ASR produces text; pause or finish before analyzing the captured recording
- 03NPU / CPU
Speaker analysis
Unload ASR and analyze speakers offline to finalize recording-local IDs and intervals
- 04CPU
Text alignment
Associate text with speaker intervals and retain uncertainty markers
- 05Browser
Replay and export
Filter speakers, replay from text or timeline segments, and export JSON
Supported hardware
- Core model parameters
- ~8.1M
- Model files
- 30.95 MB
- License
- Pandora License