EchoSync runs the full voice pipeline at the edge — capture, wake, and transcribe never leave your host. Only text travels to the LLM. The result: sub-second conversational AI with enterprise-grade privacy.
No sign-up. No account. Mic in <12 ms, reply out in under 800 ms.
Each stage overlaps the next so latency stays conversational, not lab-like.
Every subsystem has an SLA, every metric is live, every byte is accounted for.
Raw microphone audio never leaves the host. VAD and transcription run locally, so ambient sound stays on your metal. Compliant by construction.
INT8-quantized CTranslate2 and ONNX graphs hold CPU load to 1.5–2.5 cores. No GPU, no dedicated instance, no cloud ingestion fees.
Only finalized transcripts are sent upstream over TLS. Gemini Flash streams tokens back while TTS already renders the first sentence.
Sentence-boundary chunking masks time-to-first-audio. Kokoro synthesizes the first clause while the model is still finishing its reasoning.
Speak mid-reply and playback stops instantly. Silero VAD detects your voice and the pipeline pivots back to listening within a few frames.
One command runs the entire stack. Docker-ready, health-probed, and instrumented for real-time observability out of the box.
An async, decoupled pipeline that keeps speech local and thought remote.
Audio worklet streams 512-sample PCM frames. Silero VAD gates speech with a 400 ms silence window — nothing else is ever buffered.
Faster-Whisper with CTranslate2 turns the spoken turn into text on the consumer CPU. Raw audio is deleted the moment it is transcribed.
Only the tiny text payload travels to Gemini over TLS. Streamed tokens are chunked into sentences and rendered by Kokoro as they arrive.
Launch the console, grant microphone access, and start a session. No account, no setup, no watermark.
Try Now →First words out in under 800 ms round-trip · Works on any modern browser