v0.8.4

• The downloaded dictation app no longer crashes on start. CI built
  whisper.cpp for its own CPU (AVX-512), so on CPUs without it (such as a
  Core Ultra 9 275HX) the app died with an illegal instruction right after
  loading the model. The 0.8.2 and 0.8.3 downloads were affected; local builds were
  not. CI now builds for portable x86-64 with AVX2.

v0.8.3

• Whisper runs on the discrete GPU on hybrid laptops. With NVIDIA
  Optimus the Intel iGPU is Vulkan's first device, and whisper.cpp took it:
  about 50 times slower than the RTX (real-time factor 2.3 instead of
  0.05). Each wake-phrase check took seconds, so "Oye Claude" was often
  missed. The dictation app now picks the first discrete GPU and logs it
  ([model] gpu: ...).
• Escape cancels a dictation without interrupting Claude. The Escape
  that cancelled a recording also reached the window in front, so it
  stopped whatever Claude was doing. While recording, the dictation app now
  takes Escape for itself (a system hotkey), then releases it.

Note (2026-09-29): the original claudetalk-dictation.exe of this release crashed at start on CPUs without AVX-512 (illegal instruction). It was replaced with the portable build from v0.8.4, which is compatible.

v0.8.2

Note (2026-09-29): the original claudetalk-dictation.exe of this release crashed at start on CPUs without AVX-512 (illegal instruction). It was replaced with the portable build from v0.8.4, which is compatible.

v0.8.1

Note (2026-09-29): the original claudetalk-dictation.exe of this release crashed at start on CPUs without AVX-512 (illegal instruction). It was replaced with the portable build from v0.8.4, which is compatible.

v0.8.0

Note (2026-09-29): the original claudetalk-dictation.exe of this release crashed at start on CPUs without AVX-512 (illegal instruction). It was replaced with the portable build from v0.8.4, which is compatible.

v0.7.1

chore: v0.7.1

claudeTalk 0.7.0 — native Rust

0.7.0 — native dictation, no Python left

• Dictation. bin\claudetalk-dictation.exe (Rust) replaces the Python
  daemon (voice-input\) with the same features and settings. Measured on
  an RTX 5070 Ti laptop:

  	Python daemon	Rust daemon
  RAM, idle	279 MB	13 MB
  RAM, model loaded	279 MB	97 MB
  VRAM	~2.1 GB	0.85 GB
  CPU at rest	4.5 % of a core	0.4 %
  Spanish WER on the benchmark	5.86 %	5.44 %

• Speech-to-text. Whisper large-v3-turbo q8_0 on whisper.cpp with Vulkan,
  instead of faster-whisper fp16 on CUDA, so any GPU works (NVIDIA, AMD or
  Intel). See native/bench/RESULTS.md.

• Wake phrase. Silero VAD filters out sound bursts that aren't a voice
  before Whisper runs.

• Chord. Detected from raw keyboard events instead of polling every 15 ms.

• Install. Nothing to install besides the plugin. The dictation app
  (~60 MB, not kept in git) and the models (~0.9 GB) download on first run.
  The dictation app comes from the GitHub release of the same version.

• Cleanup. claudetalk.exe cleanup [--yes] removes what older versions
  left behind. The SessionStart hook tells Claude while leftovers exist.

• Fix. "Turn off dictation" from the gear panel now also forgets the
  manual-launch flag, as the tray's button did.

• Removed: voice-input\, setup-voice.ps1, dictation.vbs,
  make-icon.py.

0.6.0 — native plugin binary

• bin\claudetalk.exe (Rust) replaces every PowerShell script: hooks, the
  say MCP server, /talk and the speech player.
• Hooks. A prompt hook takes ~17 ms instead of 0.4–1.5 s.
• Speech. Edge TTS is spoken natively. Playback starts with the first
  chunk, and back-to-back phrases start in ~20 ms. edge-tts and ffmpeg are no
  longer needed.
• Talk mode tokens. The full rules go on the first prompt and every 8th
  one; a one-line reminder goes on the rest.

---

Upgrading from v0.5 or earlier: run bin\claudetalk.exe cleanup to see what the old Python version left behind (often several GB), then cleanup --yes.