Talkback

A browser voice agent that measures itself — honestly.

checking backend…

“Talkback” is the studio channel an audio engineer uses to speak to the artist — I ran one for 8 years before I started shipping agent systems. This console is a working reverse-application: point it at your brand in minutes.

Voice Console

PREVIEW · agent hop simulated & labeled

IDLEPress the mic and just talk.

No turns yet. Press the mic, or tap a suggestion below — every hop gets timed either way.

Speech recognition runs in your browser (Web Speech API). Best in Chrome, Edge or Safari.

Latency Observatory

budget 800 ms · user stops talking → agent starts speaking

— ms

E2E median (p50)

— ms

E2E p95

turns measured

LAST TURN vs BUDGET—

STT finalize network inference simulated agent hop (preview) TTS queue

Where this sits in a production voice stack

This demo — every hop measured, in your browser, right now

Browser micgetUserMedia

STTWeb Speech API · measured

AgentWorkers AI · net + inference split

TTSspeechSynthesis · measured

Speakere2e vs 800 ms budget

Production — your stack: same budget, different transports

PSTN / SIP trunkcall arrives

RTP mediaaudio frames

Streaming STTpartials, not finals

Agent coreorchestrated, autoscaled

Streaming TTSfirst-byte matters

Honest scoping: I built the top row and the discipline of measuring it. The bottom row — SIP, RTP, the streaming pipeline — is your team’s craft, and it’s the part I’d be learning from day one. What transfers is the part this page demonstrates: shipping a working system fast, instrumenting every hop, and refusing to quote a latency number I didn’t measure.