Pulse
Voice inference gateway with cancellation and recovery
Connects speech recognition, language-model inference and speech synthesis through an authenticated WebSocket gateway — faster-whisper, SmolLM2 and Piper behind bounded admission, circuit breakers, fallback before audio starts, interruption handling, worker recovery and latency instrumentation.
MeasuredFirst-audio p99 of 734.14 ms at concurrency 4 on local CPU (8-token output cap, 8 samples), with worker-kill recovery and stale audio rejected after interruption. 17 tests pass.
Local CPU inference and recorded demo; continuous microphone capture, hosted inference and GPU measurements are unfinished.