v0.1.0-alpha.3 — 2026-08-06

WHAT'S NEW IN v0.1.0-alpha.3 (2026-08-06)

NEW THIS BUILD (since the 2026-07-23 build)

SPEECH RECOGNITION NOW RUNS ON YOUR GPU

• The big one this build. Local speech recognition now runs on your
  graphics card when a usable one is there. Measured 4-19x faster
  recognition on a discrete GPU, and about 4x even on integrated
  graphics - because recognition stops competing with the game for
  CPU cores.
• It picks your discrete GPU on its own (integrated when that's all
  there is). A machine without a usable Vulkan GPU stays on the CPU
  engine exactly as before, and the log says so in one line.
• Advanced, only if your rig disagrees: create appsettings.local.json
  next to TeaTools.App.exe and set Stt:Backend (auto / cpu /
  vulkan), Stt:GpuDevice (device number), or Stt:WhisperThreads
  (how many CPU cores decoding may use). Leave them unset and
  TeaTools decides.

THE DOWNLOAD GOT BIGGER: ABOUT 53 MB -> ABOUT 80 MB

• Worth knowing before you click download. The GPU engine ships
  inside every build, whether or not your machine can use it, and
  that is most of the extra ~27 MB (~20 MB of GPU engine, ~7 MB of
  the v2 shell + IBM Plex Sans JP typeface). The alpha.2 zip was
  about 53 MB; this one is about 80 MB.
• Nothing you download separately changed: the free local speech
  model is still a one-time ~142 MB download on first listen, cached
  forever after.

FIXED: VRCHAT STUTTER WHILE TRANSCRIBING

• The GPU backend above is the main fix now - recognition mostly
  moves off the CPU entirely, so it stops fighting VRChat for it.
• On CPU-only machines the original cause is fixed too: speech
  recognition was asking Windows for EVERY CPU core, and on many
  machines transcription couldn't even keep up with speech while
  VRChat fought it for the CPU. It now uses a small, bounded share.
  That bound still applies on the GPU path, to the decoding work that
  stays on the CPU.

THE APP HAS A NEW LOOK (v2 SHELL)

• The whole window is reorganized: modules live in a rail on the left,
  app settings moved into Home, and the window chrome, dialogs and
  tooltips were rebuilt on the new design system. Everything you had
  is still there - it may just be in a clearer place now.
• New app typeface (IBM Plex Sans JP). Japanese text keeps full
  coverage.

FACE TRACKING + TEATOOLS AT THE SAME TIME

• TeaTools no longer fights VRCFT (or any second OSC app) for
  VRChat's ports. It binds its own port, announces itself properly,
  and VRChat routes to both apps. Verified live with VRCFT + OVR
  Toolkit running together.

MUTE SAFETY HARDENING

• After a VRChat reconnect, TeaTools no longer trusts stale "you're
  unmuted" state from before the drop - it waits until VRChat
  reports your real mute state. It also re-checks that state in the
  background instead of trusting one reading forever, and stops
  trusting a reading once it has gone stale.
• NEW WARNING PILL: once TeaTools has connected to VRChat, if it later
  can't verify your mute state AND something is being held back
  because of it, a warning appears under the title bar - "Mute state
  unknown - paused to be safe. Resumes once VRChat reports you're
  unmuted." - with a Check OSC button that jumps to Home -> General.
  It never auto-dismisses, retracts itself the moment the state is
  verified again, and never fires before that first connection - there's
  nothing to warn about yet, so TeaTalk behaves normally until then.
  Before this pill, held speech was silent - indistinguishable from the
  app just not hearing you. It now covers TeaTalk itself, not just
  third-party modules: from that first connection onward, whenever
  TeaTools can't tell whether you're muted - including a later
  disconnect - it holds your speech instead of guessing, and tells you
  it's doing so.
• Third-party modules without the "hear while muted" permission now
  receive speech only when TeaTools has PROVEN you were unmuted at the
  moment you started speaking. Before, in the window between app start
  and VRChat's first mute report, they were treated as if you were
  unmuted - speech made while already muted could reach them. Unknown
  now means withheld, from your first VRChat connection onward.
• CLOSED: the last gap in this hardening. Previously, once connected,
  a later OSC disconnect could make TeaTalk's own chatbox sends
  resolve an unverifiable mute state as "not muted" - the one path
  where an unknown state could still be sent instead of held. Unknown
  now means held on every spoken path for the rest of that session,
  however the connection behaves.

FIXED: THE X BUTTON NOW ACTUALLY CLOSES TEATOOLS

• With "close to tray" turned off (the default), clicking X used to
  make the window disappear while the app kept running in the
  background - mic still open, still sending to the VRChat chatbox,
  nothing on screen to show it. The only way out was Task Manager. X
  now actually exits the app. If you've turned "close to tray" on,
  that's unaffected - X still hides to the tray as before, and you
  quit from the tray icon's Quit.

NEW: TRANSPARENCY EFFECTS SWITCH (ACCESSIBILITY)

• If text is hard to read through the frosted-glass look, you can now
  turn it off: Home -> General -> Appearance -> "Transparency
  effects". The window goes fully opaque - same design, nothing
  showing through. Changes apply instantly, no restart.
• Known issue: in the LIGHT theme, some muted grey text is still
  lower-contrast than we want even with transparency off - an ink
  color fix is tracked separately.

SMALLER BUT REAL

• About now has working Website and Discord links (and the broken
  "Source code" link is gone).
• The OSC status dot now tells you HOW to fix it - hover it when
  disconnected and it names the VRChat menu path to re-enable OSC.
• Clearer words in a few places: "Pause transcribing" (was "Pause
  STT"), "modules" everywhere (was a mix with "plugins"), and the
  free-path banner no longer overclaims.
• Testers' console windows no longer fill up with internal
  diagnostic lines.

Coming from a build before 2026-07-23? You also have: Live captions
(opt-in preview), the repeated/doubled-words fix, and the
mute-decided-at-speech-onset privacy fix.

TeaTalk v0.1.0-alpha.2

Alpha build for trusted testers — Windows x64, self-contained (no install; unzip and run TeaTools.App.exe).

What's new since the 2026-07-17 build

💬 Live captions (opt-in preview)

TeaTalk can now show your words in the chatbox as you speak, before the finished translation arrives. OFF by default — try it via TeaTalk settings → Chatbox → "Live captions". Flipping it briefly pauses listening while the pipeline restarts; that's normal. It's a preview: if anything feels off, flip it back off and tell us what you saw.

🔧 Fixed: repeated/doubled words

With Live captions on, longer sentences could come back with words repeated or doubled before translation — garbling what got sent. Root-caused and fixed; verified in a live session before this release.

🔒 Privacy fix: mute is decided the moment you speak

In a rare timing window, speech that started while your VRChat mic was muted could still reach the chatbox with mute set to Pause STT. The mute decision is now locked at the instant you start speaking and can't be raced.

🔍 Under the hood

Much sharper latency + diagnostic logging (per-stage timings in %AppData%\TeaTools\logs). When something feels slow, the log now shows exactly which stage did it — and if you opted into crash reporting on first run, much of that reaches us automatically.

---

Coming from a build before 2026-07-17? You also have: typed messages send even while your mic is muted, the OSC port-in-use startup crash fix, better "it hears nothing" microphone diagnostics, and small-window layout fixes.

Full details ship in the zip (RELEASE-NOTES.txt + docs/). Rough edges expected — tell us what breaks. 🙏

TeaTalk v0.1.0-alpha.1

Say it once. They read it in their language.

Real-time speech translation for VRChat. You talk normally; the people around you read it in their language, right in the in-game chatbox — while you keep talking.

Free to use · No account · Windows

What it feels like

You join a world, someone says hi in a language you don't speak, and you just... answer. You speak normally into your mic. A beat later, your words land in the VRChat chatbox in their language — no menus, no stopping mid-sentence, no copy-pasting into another window. Flip to transcribe mode and it captions your own voice instead, clean and live.

It works the first time you open it. Nothing to sign up for, nothing to pay.

Why you'll like it

• 🗣️ Speak, don't type. Your voice becomes chatbox text in another language while you keep talking.
• 🆓 Free out of the box. Speech recognition runs on your PC (local Whisper); translation is built in (Microsoft Edge, with automatic fallback to Google). No account, no API key.
• 🌐 Auto-detects your language. Just talk — TeaTalk figures out what you're speaking. Prefer to pin one language? Pick it, and local recognition gets faster.
• 🔀 Translate or transcribe. Post your words in another language, or caption your own speech as-is. One click to switch.
• 📝 Reads the way you want. Show the original first, the translation first, or translation only.
• 🔇 Mute players are first-class here. Keep your VRChat mic muted, switch on Ignore mute, and TeaTalk still turns your speech into chatbox text — so you're in the conversation while your voice stays off.
• 🧠 Optional: bring your own key. Point TeaTalk at your own OpenRouter model for conversation-aware translation — the free path needs no key.

Download

1. Download TeaTools-alpha.zip below and unzip anywhere (your Desktop is fine).
2. Run the app — nothing else to install.
3. In VRChat, enable OSC: Action Menu (R) → Options → OSC → Enabled. TeaTalk finds VRChat automatically once OSC is on.

First launch: Windows may show a SmartScreen notice because the app isn't code-signed yet — click More info → Run anyway. (Alpha; signing comes later.)

You'll need: Windows (64-bit) · VRChat with OSC enabled · a microphone.

Build refreshed 2026-07-15 (0502e0b): adds internal speech-timing diagnostics so tester reports pinpoint transcription lag. No user-facing feature changes since the initial build.

What's new in v0.1.0-alpha.1

• Auto-detect is the default. Fresh installs detect your spoken language automatically and translate each phrase from what you actually spoke, on every engine. There's a new "Detect language" option at the top of the picker. Already set a language? Nothing changes — and picking one is still the faster path.
• Faster, more accurate recognition when you pick your language. Setting your language locks local speech-to-text onto it instead of guessing every phrase; changing it mid-session (including pressing Swap) re-locks right away. This now works for every language in the picker, and uses all your CPU cores.
• Respect VRChat mute now works even if you were already muted when TeaTalk started — it reads your current mute state from VRChat the moment it connects and holds your speech until it knows. Muted means muted.
• Reliability: speech recognition survives engine hiccups (a cloud timeout no longer silently kills transcription for the session); unplugging your mic stops listening honestly instead of pretending it's still on; and the free translator recovers faster during an outage.
• Honest diagnostics wording. Crash/problem uploads were called "anonymous" — they're not: no account or identifier is attached, but logs can include file paths from your PC. The wording now says so plainly, and crash reports strip your Windows username from those paths.
• Overhearing others is strictly opt-in — turning it off stops the listening itself, not just the messages.
• Safer builds and a batch of concurrency/shutdown edge cases hardened out.

Known rough edge: auto-detect judges each phrase on its own, so very short or noisy clips can occasionally be detected as the wrong language. If that bites you, pick your language in TeaTalk — it's also faster.

How it works (and what's free)

The default path is free and runs on your machine: speech becomes text on your PC with local Whisper (your voice audio isn't sent to a server on this path), and translation uses free, keyless engines (Microsoft Edge, falling back to Google automatically). No account, no key, no catch.

Optional power-user mode: add your own OpenRouter key and translation works like a conversation — conversation memory carries context between lines, a glossary keeps names/jargon consistent, and links and @mentions survive intact. Any keys you add are encrypted on your PC (Windows DPAPI) and never leave it. You never need a key to use TeaTalk.

Provider names (Whisper, Microsoft Edge, Google, OpenRouter, VRChat) describe what TeaTalk connects to — not endorsements.

This is an alpha 🌱

TeaTalk is early and built in the open. Expect rough edges, expect things to change. Built in the open on Discord — the feedback you leave now steers what ships. If something breaks, the in-app "Report a problem" button is the fastest way to help fix it.

License

TeaTalk is part of TeaTools — released under PolyForm Noncommercial 1.0.0 (free to use, not open-source-for-commercial-use).