Skip to content

hotfix: 1.62.1 — stop a phone's own voice coming back to it, live spend for the host - #442

Merged
micio86dev merged 6 commits into
mainfrom
hotfix/1.62.1
Nov 20, 2026
Merged

micio86dev merged 6 commits into
mainfrom
hotfix/1.62.1

Conversation

@micio86dev

Copy link
Copy Markdown
Owner

Reported from a real two-phone test of Live Connect: the owner (Italian) spoke next to the partner (Spanish); the partner got the correct translation but the owner ALSO got their own sentence back in Italian, because their voice reached the partner's microphone and was "translated" by the partner's engine session.

Changes

  • Echo filter (server/src/live/echo.rs, wired into each seat's outbound pump): drops a partner utterance whose ASR original resembles my recent own speech, only when my microphone was clearly louder; holds a new utterance at most 700 ms and only when I spoke in the last ~12 s; fails open. Kill switch LIVE_ECHO_FILTER=0, tunable LIVE_ECHO_SIM. N-seat ready in the state machine.
  • Live spend for the host: session_spent frame (owner only) and a host line Credits left: $X.XXXX · −$Y.YYYY to four decimals, painted from the cached balance at start.
  • 19 Live Connect strings corrected in low-resource locales after an independent review.
  • Hardening after an independent verification: fast-attack mic level, an unknown level never suppresses, a lone word is never judged, bounded text.

Verification

  • Server: 2214 tests pass (only the two known unrelated voip::regulatory flakes), clippy -D warnings clean on 1.97 and 1.98.1; 20/20 mutations of the filter killed; verifier probes adopted as tests/live_echo.rs. Client: 2559 tests, astro check 0 errors, build ok.
  • Deployed to staging first: healthy, no errors, filter metrics present.

Not verified

Real Qwen ASR behaviour on cross-language echo and real acoustics: only the mock engine. First signals if echo persists: voxtranslate_live_echo_failopen_total vs ..._suppressed_total, then lower LIVE_ECHO_SIM toward 0.5. Echo utterances are still metered and persisted (the filter only stops delivery).

When two phones sit side by side, one phone's microphone also hears the
other's speaker, so the engine "translates" that speech and delivers it
back to the person who said it. Each seat's outbound path now passes an
EchoFilter (live/echo.rs, pure state machine, N-seat ready):

- ledger of every speaker's recent original words, tagged by speaker;
- an utterance of speaker X is an echo when its ASR `original` (never the
  translated `text`) is similar to a recent entry of somebody else AND that
  somebody's microphone is clearly the louder one (prevents both seats
  dropping each other's real speech);
- a new utterance is held at most 700 ms only when somebody spoke
  recently, released in order if real, dropped whole if echo, and fails
  open on timeout; bounded buffers and speaker table;
- LIVE_ECHO_FILTER (default on) and LIVE_ECHO_SIM (0.3-0.95, default 0.6);
- metrics live_echo_suppressed_total / live_echo_failopen_total and one
  diagnostics event per conversation per 10 s (counts only).

Persistence and billing are untouched. Logic, state machine and the mock
engine path are tested; real ASR/acoustics are not verified.
…l never suppresses, lone words never judged, bounded text) + adopted verifier probes
@vercel

vercel Bot commented Nov 20, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
voxtranslate Ready Ready Preview Nov 20, 2026 11:44am UTC

@micio86dev
micio86dev merged commit ce9c937 into main Nov 20, 2026
11 checks passed
@micio86dev
micio86dev deleted the hotfix/1.62.1 branch November 20, 2026 12:18

This branch was successfully deployed

1 active deployment
Preview — 43a2a022 Deployed Nov 20, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant