v0.2.0 Grok Bot & Live visualization

Persona Voice now changes the output voice of Grok Bot, alongside ChatGPT and Codex. Here it is in a live conversation:

Voice-changer-Grok-demo-40s-v3.mp4

What's new

• Select Grok Bot or ChatGPT / Codex as the source application.
• Compare original and converted speech in the new Live waveform view, with model processing time and installed package size.
• Grok Bot and Codex demos are now side by side in the English, Japanese, and Chinese READMEs.
• Updates now use the renamed miuuyy/persona-voice repository.

Tested

Manually tested by the maintainer in live voice conversations on a MacBook M4 Pro; voice conversion and the updated interface worked as expected. Automated source, native-helper, and packaging checks passed on macOS, Windows, and Linux. These checks do not replace live Windows/Linux GPU testing.

Downloads

macOS Apple Silicon: .dmg or .zip · Windows x64: .exe · Linux x64: .AppImage.

Voice models download separately during setup. Windows requires VB-CABLE. Chatterbox uses Apple MLX; Seed-VC also supports the documented Windows/Linux NVIDIA profiles. OS signing/notarization limitations remain unchanged.

Upgrading from 0.1.x: download this release manually if the old updater rejects the renamed repository's asset URLs. Existing settings and models are retained.

v0.1.6

ChatGPT Persona Voice v0.1.6 adds Chatterbox as a second local voice model for Apple Silicon.

Added

• Adds the Chatterbox streaming voice model through Apple MLX.
• Adds first-run model selection: download one model during onboarding and add the other later in Settings → Voice model.
• Keeps existing installations on their selected model and retains Seed-VC on supported Windows/Linux profiles.
• Keeps the bounded route and standby handoff behavior fail-closed.

All source tests and target-native package gates passed. Chatterbox model-free streaming, installer, worker, model-selection, native, and packaged renderer checks passed. Apple Silicon live English listening is accepted. Production signing/notarization, clean-machine qualification, and physical Windows relay confirmation are still requested.

v0.1.5

ChatGPT Persona Voice v0.1.5 fixes VB-CABLE Pack45 detection on Windows 11.

Fixed

• Detects the official base VB-CABLE endpoint using the stable driver identity CABLE Input + VBAudioVACWDM.
• No longer depends on friendly-name properties that are absent on the reported Pack45 host.
• Keeps the lossless standby/conversion handoff introduced in v0.1.3 unchanged.
• Fixes macOS engine installation when the app path contains spaces, including the default /Applications/Codex Persona Voice.app location.

Thanks to @ForceofwiII for the physical Windows evidence and original driver-identity implementation.

All source tests and target-native package gates passed. The Windows helper and compiled identity regression were built with MSVC, and the generated NSIS package was installed, launched, verified, and removed successfully. Physical Windows relay confirmation is still requested.

v0.1.4

ChatGPT Persona Voice v0.1.4 fixes VB-CABLE Pack45 detection on Windows 11.

Fixed

• Detects the official base VB-CABLE endpoint using the stable driver identity CABLE Input + VBAudioVACWDM.
• No longer depends on friendly-name properties that are absent on the reported Pack45 host.
• Keeps the lossless standby/conversion handoff introduced in v0.1.3 unchanged.

Thanks to @ForceofwiII for the physical Windows evidence and original driver-identity implementation.

All source tests and target-native package gates passed. The Windows helper and compiled identity regression were built with MSVC, and the generated NSIS package was installed, launched, verified, and removed successfully. Physical Windows relay confirmation is still requested.

v0.1.3

demo.mp4

ChatGPT Persona Voice replaces ChatGPT and Codex voice output locally with the voice you choose. v0.1.3 fixes the Windows Pack45 readiness and audio-stream handoff failures reported during the first physical-host test.

Fixed

• Recognizes the official base VB-CABLE Pack45 endpoint on Windows 11 installations where the adapter-interface property is not projected.
• Waits for native self-test stdout/stderr to close before validating the final CPV1 readiness frame, eliminating false code=0 failures.
• Keeps one WASAPI capture consumer alive across standby and conversion, buffering a bounded handoff until Seed-VC and output are ready instead of dropping frames or resetting sequence truth.
• Preserves frames in both Start and Stop transitions; real in-stream sequence gaps remain terminal.

Thanks to @ForceofwiII for the Pack45 reproduction, detailed diagnostics, and physical Windows evidence.

Downloads

• macOS 14.2+ / Apple Silicon: DMG or ZIP
• Windows x64 + NVIDIA, build 20348+: EXE
• Linux x64 + NVIDIA: AppImage

Windows requires VB-CABLE, installed separately from the official VB-Audio site. The model runtime is installed separately from inside the app and is not bundled into the launcher download.

Verification

All 188 source tests, typecheck, renderer builds, native builds, and target packaging gates passed. The Windows job compiled and executed the Pack45 identity contract, built the NSIS installer, installed it into a clean temporary profile, launched its packaged renderer, and removed it successfully. Physical Windows + VB-CABLE + NVIDIA relay retesting is still requested.

Voice conversion currently performs best with Japanese and Chinese source speech. English works, but pronunciation and timbre consistency can vary. Contributions improving multilingual quality are welcome.

v0.1.2

demo.mp4

ChatGPT Persona Voice replaces ChatGPT and Codex voice output locally with the voice you choose. v0.1.2 is a startup hotfix for the macOS, Windows, and Linux preview packages.

Fixed

• Fixed the packaged app failing at startup with ERR_FILE_NOT_FOUND while loading its interface.
• Replaced privileged local-file page loading with a confined application protocol while keeping Electron's hardened file-protocol fuse disabled.
• Added release gates that launch the packaged macOS app and install, launch, verify, and remove the generated Windows installer before publication.

Downloads

• macOS 14.2+ / Apple Silicon: DMG or ZIP
• Windows x64 + NVIDIA, build 20348+: EXE
• Linux x64 + NVIDIA: AppImage

Windows requires VB-CABLE, installed separately from the official VB-Audio site. The app guides the download, restart, Volume Mixer assignment, and live route verification; VB-CABLE is not bundled.

The model runtime is installed separately from inside the app and is not bundled into the launcher download.

Verification

The exact v0.1.1 failure was reproduced from the published package before the fix. The v0.1.2 macOS package then mounted its interface successfully with the hardened Electron fuses still active. On the Windows runner, the generated NSIS package was installed into a clean temporary directory, the installed EXE mounted the complete interface, and the package was removed successfully.

Physical Windows + VB-CABLE + NVIDIA relay feedback is still requested for this preview.

Voice conversion currently performs best with Japanese and Chinese source speech. English works, but pronunciation and timbre consistency can vary. Contributions improving multilingual quality are welcome.

v0.1.1

demo.mp4

ChatGPT Persona Voice replaces ChatGPT and Codex voice output locally with the voice you choose. v0.1.1 adds the first Windows preview while keeping the same local Seed-VC relay, voice catalog, multilingual launcher, and private history controls.

Downloads

• macOS 14.2+ / Apple Silicon: DMG or ZIP
• Windows x64 + NVIDIA, build 20348+: EXE
• Linux x64 + NVIDIA: AppImage

Windows requires VB-CABLE, installed separately from the official VB-Audio site. The app guides the download, restart, Volume Mixer assignment, and live route verification; VB-CABLE is not bundled.

The model runtime is installed separately from inside the app and is not bundled into the launcher download.

Tested in real use

Manually tested end to end in live, multi-turn ChatGPT voice conversations on a MacBook Pro with M4 Pro. Voice replacement, repeated replies, voice switching, interruptions, local playback, and the final OBS demo recording were exercised in the real app flow and worked correctly during the acceptance session.

This is hands-on macOS acceptance, not a universal latency benchmark or a claim that every hardware and audio topology has been qualified. The Windows user-mode path, installer, and native helpers pass target-native CI; physical Windows + VB-CABLE + NVIDIA feedback is requested for this preview.

Included

• Near-real-time local Seed-VC conversion with 300 ms input blocks and streamed output.
• Native process-audio routing for macOS and Linux.
• Windows x64 preview with WASAPI process capture, VB-CABLE route verification, standby passthrough, and an NSIS installer.
• Guided VB-CABLE setup in English, Japanese, and Simplified Chinese.
• Bundled credited voice references, integrity checks, and per-voice character scenes.
• Optional converted-only history with retention controls.
• Signed release manifest, checksums, and pinned Bun 1.3.14 / uv 0.11.14 bootstrap runtimes.

Voice conversion currently performs best with Japanese and Chinese source speech. English works, but pronunciation and timbre consistency can vary. Contributions improving multilingual quality are welcome.

v0.1.0

demo.mp4

ChatGPT Persona Voice replaces ChatGPT and Codex voice output locally with the voice you choose. This first public preview ships the guarded native audio relay, pinned Seed-VC realtime profile, local voice catalog, multilingual launcher, and private history controls.

Downloads

• macOS 14.2+ / Apple Silicon: DMG or ZIP
• Linux x64 + NVIDIA: AppImage
• Windows x64 + NVIDIA: source is included and CI-compiled; a prebuilt installer is withheld until the virtual audio driver receives Microsoft kernel-policy signing.

The model runtime is installed separately from inside the app and is not bundled into the launcher download.

Tested in real use

Manually tested end to end in live, multi-turn ChatGPT voice conversations on a MacBook Pro with M4 Pro. Voice replacement, repeated replies, voice switching, interruptions, local playback, and the final OBS demo recording were exercised in the real app flow and worked correctly during the acceptance session.

This is hands-on macOS acceptance, not a universal latency benchmark or a claim that every hardware and audio topology has been qualified.

Included

• Near-real-time local Seed-VC conversion with 300 ms input blocks and streamed output.
• Native process-audio routing for macOS and Linux, with Windows implementation available from source.
• English, Japanese, and Simplified Chinese interface localization.
• Bundled credited voice references, integrity checks, and per-voice character scenes.
• Optional converted-only history with retention controls.
• Signed release manifest, checksums, and pinned Bun 1.3.14 / uv 0.11.14 bootstrap runtimes.

Voice conversion currently performs best with Japanese and Chinese source speech. English works, but pronunciation and timbre consistency can vary. Contributions improving multilingual quality are welcome.