24/7 real-time streaming transcription
Core model parameters ~4B
Model files 2.30 GB
Text appears as people speak; hours on end with flat speed and memory
- Fast response
- Unlimited length
- Chinese & English
Recorded on the board
Screen recording of the board web demo, recorded on Qualcomm QCS8550.
Highlights
Live transcription
Text comes out while the audio is still playing, less than a second behind the speaker.
Unlimited length
Transcribe hours of audio; memory and speed stay constant.
Knows when a speaker has finished
The model predicts the end of each turn, so paragraphs break where the speaker actually stops.
Fully on-device
Everything runs locally on the board's NPU: no cloud, no GPU.
Chinese and English
One SDK; the models for each language load in about 2 s.
Easy to integrate
A small C API and a Python binding; feed audio in any chunk size.
Pipeline
From input to output: where every step runs.
- 01Browser
Audio in
Video / mic → 16 kHz PCM16, 100 ms chunks
- 02CPU
Silence gate
Silero VAD, mutes applause and noise before the first words
- 03NPU
Streaming ASR
Speech encoder + language model, one step every 160 ms
- 04NPU
Semantic VAD
End-of-turn probability for 0.5/1/2/3 s with every token
- 05Browser
Text out
Server-Sent Events, incremental UTF-8 decode, paragraphs
Supported hardware
- Core model parameters
- ~4B
- Model files
- 2.30 GB
- License
- Pandora License