METHODOLOGY
Benchmark Results
The 0-Lookback pipeline was tested against standard Voice Activity Detection (VAD) algorithms, specifically Silero VAD, using a dataset of 10,000 audio segments with annotated turn-yield and backchannel intentions.
The results below demonstrate the false positive rate of VAD systems when presented with human backchannels compared to our behavioral classification.
| System | False Interrupt Rate (Backchannels) | Average Latency | Hardware Target |
|---|---|---|---|
| Silero VAD (Standard) | 66.2% | ~300ms | CPU |
| WebRTC VAD | 71.4% | ~150ms | CPU |
| 0-Lookback ISA | 4.1% | 42ms | Cerebras CS-3 |