Skip to content

release: 1.63.0 — several guests per link, tier choice and Enhanced with the host's cloned voice - #445

Merged
micio86dev merged 20 commits into
mainfrom
release/1.63.0
Nov 20, 2026
Merged

micio86dev merged 20 commits into
mainfrom
release/1.63.0

Conversation

@micio86dev

Copy link
Copy Markdown
Owner

Release 1.63.0 of Live Connect, merged to develop in #444 with green CI and verified on staging (migrations 069/070 applied, no errors, tier endpoint live).

  • Several guests on one link (host + up to 3, any languages; the host pays every seat per distinct target language).
  • Tier choice for the host: Standard / Premium (Gemini) with whole-session fallback; Enhanced (Cartesia) with the host's cloned voice, opt-in per conversation with explicit consent.
  • Hardening after an independent verification (one opening charge per seat per interval, narrower Enhanced echo guard, host-only fallback, bounded translation cost).
  • 38 new strings in 83 locales (model translation, not native review).

Deploy note: the v1.62.1 binary refuses to boot against a database migrated with 069/070 (VersionMissing), so roll the schema forward with the binary. Not verified with real Cartesia, Gemini, Qwen, Groq or real phones.

micio86dev and others added 20 commits November 20, 2026 09:00
chore: back-merge release 1.62.0 into develop
chore: back-merge hotfix 1.62.1 into develop
A Live Connect link used to admit exactly one guest. The host can now ask for
`max_guests` 1-3 (default 1, `LIVE_MAX_GUESTS` lowers the ceiling) and the same
link/QR admits guests one after the other, languages free.

- migration 069: `live_sessions.max_guests` (guarded NOT VALID check); the
  guest-side unique index becomes a plain index under the same name so that
  re-running 065 stays a no-op. The owner-side unique index stays.
- join_guest counts guests under FOR UPDATE on the session row, so N
  simultaneous joins admit exactly `max_guests`; `session_full` past that, the
  join window only closes a session nobody joined, and a join the engine cannot
  serve (each guest adds streams for every speaker) is refused `live_busy`
  inside the same transaction.
- a guest hanging up in a multi-guest session leaves instead of ending it.
- billing: every phone bills one stream per distinct target language, always,
  whatever the engine metadata says (the scale flag is gone from LiveMeter).
- echo filter and latency probe: per-seat (4) mic levels, last-voiced and
  stamps; the pump resolves a speaker id to its seat; one utterance is one
  latency observation however many phones hear it.
- dashboard list/detail: first guest plus `guest_count` instead of a scalar
  subquery that would 500 with several guests.
- docs, spec and .env.example updated.
The host chooses 1-3 guests at setup (max_guests); the session machine,
captions, partners panel, direction line and invitation panel now work with
several partners, and a guest leaving a multi-guest conversation sees a
"you left" screen without ending it for the others. 18 new i18n keys are
copied in English to every locale until they are translated.
…andard

The host chooses the tier when creating a Live Connect conversation
(`engine`: standard by default, or gemini_live_translate). Every seat uses it
and the owner pays every seat at that tier's rate per distinct target language.

- create validates against the registry: engine_not_available,
  engine_not_supported_yet (Enhanced, client-direct), engine_not_allowed
  (business stays Standard); the tier is stored in live_sessions.engine_id and
  returned by create, join and GET session
- GET /api/live/engines lists the selectable tiers with rate, minimum balance
  and a capacity hint
- /ws/live resolves the engine per start from the conversation, not the
  default; the meter bills at that engine's rate
- minimum balance scales with the tier's rate (same runway as Standard's
  minimum), at create and at every start
- Premium at capacity moves the whole conversation to Standard: shared
  override in LiveState, stored engine_id updated, one engine_downgraded frame,
  a diagnostics event, and Standard's rate afterwards; never live_busy
- admin transcript view exposes the engine id; docs and spec updated
Enhanced (cartesia) is client-direct, so each phone runs Cartesia STT on its
own microphone and sends text, the server translates once per distinct other
language and addresses every listener, and each listener's phone speaks the
translation with Cartesia TTS.

- create accepts engine "cartesia" (consumer hosts, scaled minimum balance)
  and voice_clone (per conversation, needs a clone and cloning enabled);
  stored as live_sessions.use_voice_clone (migration 070)
- POST /api/live/sessions/{id}/enhanced/session mints the Meet-shaped token
  for a host or a guest, rate limited per seat, conversation and address
- /ws/live: translate_text, enhanced_speak, enhanced_fallback; the owner pays
  every seat at the Enhanced rate for accepted frames; a text-level echo
  guard; the fallback moves the whole conversation to Standard
- GET /api/live/engines lists Enhanced with the user's voice clone state
Radio cards fed by GET /api/live/engines (Standard only while it loads or fails),
remembered per browser and re-validated, disabled below the tier's minimum balance.
The room shows the tier as a badge and a polite, dismissible notice on
engine_downgraded; new create errors have friendly messages. The stored session
gains an optional engine.
…nnect-multi-guest-tiers

# Conflicts:
#	docs/live-connect.md
Live Connect gets the client half of Enhanced (Cartesia, client-direct):
each phone runs Cartesia STT on its own microphone and sends text, the
server translates, and each listener's phone speaks the translation with
Cartesia TTS, in the host's cloned voice when the host opted in.

- enhanced.ts: token before start and refreshed before expiry, STT on the
  phone's stream, partial coalescing (<= 5/s), stable segment ids, bounded
  TTS queue in arrival order, voice_id else default voice, every failure
  ending in enhanced_fallback / not_enhanced / gone
- conversation: Enhanced vs PCM by session engine; engine_downgraded (or a
  409) switches to PCM; the echo guard gates the recogniser while the
  phone speaks; full cleanup on mute/end/release/reconnect
- picker: Enhanced card, hidden on restricted networks; host consent
  checkbox for the cloned voice and inline "Prepare my voice"
- room: Enhanced badge, "Speaking" line, Enhanced-specific downgrade
  notice, a guest line about the host's imitated voice
- shared with Meet without changing it: toCartesiaSession, ttsRequest,
  parseTtsFrame, openSttGraph/closeSttGraph (cartesia.ts) and
  recordVoiceSample (moved verbatim from app.ts to voice-sample.ts)
- 9 new i18n keys in all 84 locales (English text)
… fallback, bound translation cost

- billing: at most one opening charge per seat per interval, so a guest's
  start/stop loop can no longer drain the owner (Standard, Premium, Enhanced;
  user and org wallets).
- enhanced: the text echo guard judges only 3+ words at >= 0.85 similarity,
  with 2.5 s / 4 s windows, and never drops a seat that was being played the
  other seat's words or that said it first; a short reply always goes through.
- enhanced_fallback is honoured from the host only; a guest gets
  fallback_not_allowed (rate limited) and nothing changes.
- translate_text cost: partials 2/s and cut to their last 400 characters, 2
  translation calls in flight per seat, a per-conversation per-minute ceiling
  (LIVE_ENHANCED_CALLS_PER_MIN, default 600), finals reserved room.
- LIVE_STATUS_CHECK_MS makes the 3 s status check testable.
- tests: a flaky Standard tick read now polls; the Enhanced meter ending on
  the status check is pinned.
…swers both phones with one sentence) and expect the new full-session message
…k, so they can never deadlock in one process
…est-tiers

feat(live): several guests per link, tier choice (Standard/Premium/Enhanced) and the host's cloned voice
@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 5:47pm UTC

@micio86dev
micio86dev merged commit 494fcac into main Nov 20, 2026
17 of 18 checks passed
@micio86dev
micio86dev deleted the release/1.63.0 branch November 20, 2026 18:27

This branch was successfully deployed

1 active deployment
Preview — d46d978d 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