Browser voice AI fails on audio routing: the echo canceller cannot see Web Audio playback

Echo cancellation is a subtraction job, and it needs a copy of the thing being subtracted. In a browser that copy arrives only from paths the browser itself owns: an `<audio>` element, or a WebRTC receiver track. A Web Audio graph does not reliably supply one ^[2]. So when TTS leaves through a buffer source node because someone wanted sample-accurate amplitude for a mouth shape, the canceller has nothing to work against. The mic hears the synthetic voice cleanly and the recognizer files it as a user turn ^[3]. No threshold downstream separates that from a person, because without a reference signal a speaker's output and a human voice look the same at the microphone ^[5].

The usual patch is text matching: hold what TTS just said and drop transcripts that resemble it. The author's fix skips that layer, sending TTS from the server as a WebRTC remote track played through an `<audio>` element, reported as 99 seconds of continuous speech over open speakers with zero false user turns ^[4]. Read what that sentence requires: a server-side media path, a remote track, and an element to consume it. Client-side synthesis and direct buffer access are what you hand back.

Convergence is the same mechanism one order down. The canceller learns only while audio is actually playing, so a session that opens in silence begins with an untrained filter and the user's first 10 to 20 seconds get suppressed along with the echo, with autoGainControl ramping up from quiet making that same window worse ^[7]. The remedy is to make noise on purpose: a greeting or sound effect during the loading screen, AGC set to false, echoCancellation left alone ^[8]. The spinner is now part of the audio stack.

The media element turns out to be mandatory in the other direction too. Chromium will not hand a remote MediaStream to Web Audio until some media element starts consuming it, so even an analysis-only graph needs a muted `<audio>` with play() called on it, plus a held reference, since Chrome will collect unreferenced nodes and stop the graph without raising an error ^[11][12].

What remains after the routing is corrected is a genuine trade. AEC suppresses and distorts near-end audio during simultaneous speech, which is why proper nouns come apart at word beginnings when a user talks over the avatar ^[9]. The mitigations run three deep and the last of them is instructing the LLM to treat its input as error-prone transcription and reach for phonetically similar terms ^[10]. Audio from another process, including a video the agent opened itself, has no reference at all and so has no fix; headphones, macOS voice isolation that unsupported systems quietly ignore, and downstream filtering are the entire toolkit ^[6].

Four of the eight traps reproduced in this excerpt come back to where the reference signal comes from or whether it has converged ^[16]. The instrumentation advice follows from that: server-side energy logging at 10 Hz, one sample every 100 ms, tells distorted audio from true silence, and a meter on the transmitted stream earns its place because a muted track and an empty room are identical on the wire ^[13][15][14].