v0.5.9-beta.7 — VoiceStudio预发布2026-11-20
VoiceStudio v0.5.9-beta.7 更新:修复 Windows 首次安装时 docopt、antlr4、dora-search 源码编译失败的问题,安装下载现已兼容 Clash、v2rayN 等系统代理。OmniVoice 现支持 pt-BR 等带区域标签的语言,已保存的设计语音无需语音转文字模型即可生成;设置页不再常驻语音库「HTTP 404」横幅,反复出现的故障也不会挤掉其他诊断报告。
性能方面,项目网格与长列表改为仅渲染可视区域,数千条记录依然流畅;窗口隐藏或最小化时不再运行轮询定时器,恢复显示时刷新一次。本地模型文件夹未变化时保存不再卸载引擎。自动故障报告现在会标明出错环节(主进程、GPU、辅助进程、任务超时或网络),排查更直接。
安装包体积进一步缩小,清理了大量无用图片、旧组件与 477 条废弃翻译。Windows 安装包未签名,需在 SmartScreen 提示中选择「更多信息 → 仍要运行」;macOS 与 Android APK 已签名,iOS 为未签名 IPA,需自签侧载。
v0.5.9-beta.6 — VoiceStudio预发布2026-11-18
VoiceStudio v0.5.9-beta.6 更新
这版主要解决老显卡的 GPU 加速问题:GTX 900/1000 系列 NVIDIA 显卡通过 CUDA 12.6 版 PyTorch 重新跑上 GPU(#4242)。新增「关闭到托盘」,关掉窗口后 VoiceStudio 和任务继续在后台运行(#4174);Docker/web 版的下载按钮也修好了(#4224)。启动体验更顺:不再每次弹「首次运行」提示,慢速冷启动也不会在 300 秒后被杀掉(#4228)。另外在文件操作、HTTP 错误处理和 CI 上做了安全加固,Windows 安装报错也更清楚(#4173–#4188、#4213、#4230)。
性能方面,项目网格和长列表改为只渲染可视区域,几千条数据也保持流畅(#4079、#4074);窗口隐藏或最小化时不再跑轮询定时器,省下空闲 CPU(#4073)。本地模型文件夹保存时,若路径没变就不再卸载引擎(#4043)。
代码层面清理了不少遗留物:退役旧版共享 UI 及其测试、删掉 12 个无用 npm 包和 477 条废弃翻译,安装包体积随之变小。失败报告现在会标明出错环节(主进程、GPU、辅助进程、GPU 任务超时或网络),不再只报「未知错误」(#4243)。
下载:Windows x64 与 ARM64(实验性)安装包未签名,SmartScreen 会拦截,需点「更多信息 → 仍要运行」;macOS 两个 DMG 已签名并公证;Linux 提供 AppImage 和 deb;Android APK 已签名,AAB 供上架 Play;iOS 为未签名 IPA,需自签侧载,非 TestFlight 版本。
v0.5.9-beta.5 — VoiceStudio预发布2026-11-16
VoiceStudio v0.5.9-beta.5 更新:新增可选的自动错误报告功能,需在隐私设置中单独开启,默认关闭,且绝不包含音频、文本、文件、姓名、凭据或日志。Android 端 Stories 与有声书支持长文稿滚动及草稿恢复;Stories 可合并相邻行的兼容说话人与音色设置;Linux ROCm 环境保留已装的 PyTorch 栈,GPU 就绪状态如实反映自检失败;手机配音导出前检查可用空间,避免 iOS 误报空间不足。
性能方面,项目网格与长列表改为仅渲染可视区域,数千条记录依然流畅;窗口隐藏或最小化时不再运行轮询定时器,恢复显示时刷新一次。此外清理了大量无用遗留代码、图片与翻译条目,安装包更精简,本地模型文件夹保存不再误卸载引擎。升级可直接覆盖安装或走设置内更新,音色、项目、设置与已装模型均保留。
v0.5.9-beta.3 — VoiceStudio预发布2026-11-16
VoiceStudio v0.5.9-beta.3 主打下载与安装的可靠性:安装过程会复用已校验的 wheel,遇到瞬时网络故障自动重试;模型下载能识别并拒绝损坏的数据分段,同时保留已完成的有效进度,中断后恢复更稳。低显存场景下,生成前会先释放闲置模型,缓解小显存 GPU 的压力。发音词典的保存条目和内联覆盖现已贯通语音、Dub、批量与声音转换各环节。配音和转写支持导出纯文本,配音还能在生成前后保存转写稿,转写导出的 SRT 时间轴真实可用,长录音按时长放宽限制。Android 下载可走合规 VPN 路由,已安装的受限模型会弹出许可确认。注意:Windows 安装包未签名,SmartScreen 会拦截,需手动放行;macOS 与 Android APK 已签名。官方点名待测小显存生成、长文件转写、受限网络安装等场景,并列出若干仍未修复的已知问题。
v0.5.9-beta.2 — VoiceStudio预发布2026-11-14
VoiceStudio v0.5.9-beta.2 更新:Windows 安装器在 Python 下载卡住时会自动切换镜像重试,取消下载也不再触发主进程崩溃;OmniVoice GGUF 新增声音设计、[laughter] 等非语言标签及解码步数滑块;Stories 与 Audiobook 在窗口白屏恢复后保持打开,重启更新功能修复;本地应用(如 AIRI)可直接获取调用 API 所需配置;配音保留各克隆音色的原始录音,中断后可从已完成片段继续。性能方面,项目网格与长列表改为按可视区域渲染,成千上万条记录依旧流畅;窗口隐藏或最小化时不再运行轮询计时器,降低空闲 CPU 占用。安装包体积进一步精简,清理了未使用的图片、旧代码与 477 条闲置翻译。升级可直接覆盖安装或走 Beta 通道内更新,音色、项目与设置均保留。
v0.5.9-beta.1 — VoiceStudio预发布2026-11-13
Download
| Platform | Download | Signing |
|---|---|---|
| Windows x64 | Installer | ⚠️ Unsigned: SmartScreen shows "Windows protected your PC"; choose More info → Run anyway |
| Windows ARM64 (experimental) | Installer | ⚠️ Unsigned: SmartScreen shows "Windows protected your PC"; choose More info → Run anyway |
| macOS Apple Silicon | DMG | ✅ Signed with Developer ID and notarized by Apple |
| macOS Intel | DMG | ✅ Signed with Developer ID and notarized by Apple |
| Linux x64 | AppImage · deb | — Linux packages are not code-signed |
| Android | APK · AAB | ✅ Signed APK; the AAB is the Play Store bundle |
| iPhone | Unsigned IPA | ⚠️ Unsigned: sign it with your own Apple ID to sideload; not a TestFlight build |
Highlights
- Voice Cloning queues takes instead of refusing a second one while rendering
- Home shows a What's new link, winter pixel snow and a Christmas-week Santa
- Audiobook MP3 exports play and download again, and full dubs survive free-text directions like "slow"
- OmniVoice GGUF falls back to the CPU when the GPU crashes, and finds its runtime in moved app folders
- Long lists stay fast, and a hidden window no longer runs timers
Voice Cloning gets a generation queue. Start a take while another renders and it waits its turn, keeping the script, voice and settings you had when you clicked; cancel one or all from the queue panel. A waiting take is skipped, with the reason, if you replace, lock, unlock or re-transcribe its voice before it starts.
Testing wanted: the queue with voice edits while takes wait; OmniVoice GGUF falling back to the CPU after a GPU crash (Pascal cards such as the GTX 1080, and Apple Silicon); and long Projects, dub project and Transcribe history lists.
Thanks: @ItzSudhan (generation queue, #4064), @b84330808 (OmniVoice GGUF runtime, #4053), @blutuxyz (LAN sharing, #4059), @hkarmoush (Colab, #4069) and @kevin9327 (SSML emoji, #4020) for code; @dimfk001git (#4071), @user555gtb (#4054), @onparr-24 (#4058), @MeinLive (#4050, #4051), @JoanSPabon (#4066, #4067) and @adluanbo (#4035) for bug reports.
Upgrading: Install over your existing app, or take the in-app update on the Beta channel; voices, projects and settings are kept.
Changed
- Long Projects, dub project and Transcribe history lists now render only the rows on screen, so they stay fast with thousands of items (#4074)
- Idle CPU: a hidden or minimised window no longer runs polling timers or elapsed-time clocks, and they refresh once when it is shown again (#4073)
- Removed unused renderer files and 477 unused translation strings, and locale:check now fails on an English key no code uses (#4047)
- Removed more unused legacy shared components, utilities and API helpers from the old shared UI folder (#4045)
- Local-model folder saves no longer unload the engine when the stored folder did not change (#4043)
- Removed more unused legacy settings, gallery and engine screens from the old shared UI folder (#4042)
- Idle polling is relaxed: the backend health probe slows when steady, cloud-sync polling pauses while the window is hidden, and the Windows parent watcher polls less often (#4041)
- Removed more unused legacy dubbing screens and hooks from the old shared UI folder (#4040)
- Release notes now open with one Download table listing every desktop and phone download with its signing, followed by the Highlights (#4037)
- Removed about 25,000 lines of unused legacy interface code, two unused font packages and the always-skipped weekly evals job; the app itself is unchanged (#2603, #4032)
Added
- Cebuano is now a dub translation target (#4058, #4072) — thanks @onparr-24!
- Home shows a What's new link, winter pixel snow, a Christmas-week Santa, two rows of contributor portraits, and an easter egg that hides the credits until the next update after you visit three contributor profiles (#4070)
- Voice Cloning has a generation queue: takes started while one is rendering wait their turn and run one after another — thanks @ItzSudhan! (#4064)
- Audiobooks and Stories have a New project button that starts an empty draft without touching saved projects (#4055)
Fixed
- OmniVoice GGUF falls back to the CPU when the GPU backend crashes, for example Vulkan on a GTX 1080 (#4050, #4072) — thanks @MeinLive!
- Full dubs no longer fail when translation fills segments with free-text directions like "slow"; unsupported ones are dropped for that segment (#4054, #4072) — thanks @user555gtb!
- Audiobook MP3 exports play and download again instead of returning 404 (#4071, #4072) — thanks @dimfk001git!
- Colab: setup cells show the failing command's output, the security-floor check judges the Mako copy Python imports, and the web UI loads through the new *.prod.r.colab.dev proxy — thanks @hkarmoush! (#4069)
- Turning on LAN sharing no longer crashes the backend with an Internal Server Error — thanks @blutuxyz! (#4059)
- The OmniVoice GGUF engine finds its bundled runtime again in installed apps instead of reporting it missing — thanks @b84330808! (#4053)
- When the operating system times out starting the engine (spawn ETIMEDOUT, seen on macOS), VoiceStudio retries twice and then says to restart the computer (#4039)
- Settings → Updates lists only your channel's releases, and Beta suggests switching to Stable when a newer stable is out (#4038)
- SSML [spell] keeps flag, ZWJ-joined and skin-tone emoji whole instead of splitting them (#4020) — thanks @kevin9327!
CI
- Docker's web build no longer fails the unused-key check when only electron/ is copied, and the iPhone CI job allows for a native-runtime cache miss (#4052)
- Removed unused scripts, the old Tauri MSI tooling and leftover Turborepo ignore lines (#4046)
- Accessibility smoke timeouts now name the requests still pending, so a flaky CI failure points at its cause (#4036)
Contributors
Thanks to everyone who contributed to this release: @debpalash, @ItzSudhan, @velixio, @b84330808, @blutuxyz, @kevin9327, @hkarmoush, @RerankerGuo
v0.5.8 — VoiceStudio2026-11-12
Download
| Platform | Download | Signing |
|---|---|---|
| Windows x64 | Installer | ⚠️ Unsigned: SmartScreen shows "Windows protected your PC"; choose More info → Run anyway |
| Windows ARM64 (experimental) | Installer | ⚠️ Unsigned: SmartScreen shows "Windows protected your PC"; choose More info → Run anyway |
| macOS Apple Silicon | DMG | ✅ Signed with Developer ID and notarized by Apple |
| macOS Intel | DMG | ✅ Signed with Developer ID and notarized by Apple |
| Linux x64 | AppImage · deb | — Linux packages are not code-signed |
| Android | APK · AAB | ✅ Signed APK; the AAB is the Play Store bundle |
| iPhone | Unsigned IPA | ⚠️ Unsigned: sign it with your own Apple ID to sideload; not a TestFlight build |
Highlights
- OmniVoice GGUF uses AMD, Intel and NVIDIA GPUs through Vulkan on Windows and Linux, and Intel Arc GPUs run OmniVoice
- Setup falls back to a smaller CPU build when PyTorch downloads fail, and the app falls back to the CPU when the GPU fails
- Windows loads models one file at a time, and an engine crash names its cause and offers Restart on CPU
- dots.tts voice cloning runs on Windows through audio.cpp, with no Python setup
- Convert can keep your delivery and takes video; Dubbing exports subtitles first and moves lines between speakers
OmniVoice now uses your AMD, Intel or NVIDIA GPU, and VoiceStudio gets much harder to break. OmniVoice GGUF runs on any Vulkan GPU on Windows and Linux, sizes its model to your card and can stay loaded between requests (opt-in, Settings → Performance), and Intel Arc GPUs run OmniVoice too. When something goes wrong, the app now keeps going: setup offers a much smaller CPU build when the PyTorch download keeps failing, retries stalled downloads, and VoiceStudio moves to the CPU with a notice when the GPU fails. On Windows, models load one file at a time to avoid a crash on first launch, and an engine crash names the library behind it and offers Restart on CPU.
This release also brings dots.tts voice cloning to Windows with no Python setup, Convert that keeps your delivery and takes video, dubbing that exports subtitles first and moves lines between speakers, a voicestudio terminal command, background updates, a new macOS installer window, and hundreds of fixes to dubbing, setup, audiobooks and the phone apps.
The new macOS installer window: drag VoiceStudio to Applications (preview built from the installer's own background and icons)
The new macOS installer window (#4001).
Updating from 0.5.7: Take the in-app update offered by 0.5.7, or install over your existing app; voices, projects and settings are kept. On first launch the local runtime updates its packages in place, including the PyTorch 2.10 security update (#3884)
Updating from a 0.5.8 beta: Nothing to do: the Beta channel rolls forward onto 0.5.8 automatically. To stop getting betas, switch to Stable in Settings → Updates → Update channel.
Running from source: The Gradio demo and model training are no longer installed by default; add the ui or train extra if you use them (#4027)
Changed
- Smaller install: Gradio and WebDataset are no longer installed by default (about 81 MB less); the Gradio demo and model training now use the ui and train extras (#4027)
- The macOS installer window has a new branded background, larger icons and a clean volume name without the processor suffix (#4001)
- The backend starts lighter and faster: speech-recognition engines are checked in a short-lived helper process instead of being loaded at startup, while broken installs are still reported as unavailable (#3924, #3946)
- The app opens with less code to load: the audio and video players load with the first clip played and returning users skip the first-run setup screens, cutting the startup bundle by about 6%, and a new offline benchmark with a CI check on bundle growth tracks startup, memory, CPU and bundle size (#3924, #3940, #3943)
- Settings → VoiceStudio account lists what to sync and your devices as compact, aligned rows with icons, a Pro badge with one Get Pro link for locked items, and device rows that show OS and when each was last active (#3937)
- The runtime setup screen fits one window: one status line, the install folder with free space, one main action (or the exact fix, such as choosing another folder when the drive is full), with language, region, Python and remote options under Options and recovery tools under More fixes (#3934)
- OmniVoice GGUF can keep its runtime loaded between requests to skip the reload before each generation, as an opt-in under Settings → Performance (#3965, #3654)
- The audio.cpp runtime updates to v0.9.4; installed Breeze-TTS-2, dots.tts and Sortformer models keep working without a re-download, and the previous runtime keeps running them until you install the update (#3129)
- Updates now download in the background and install when you quit, with a compact title-bar pill for progress, Restart and Retry, clearer error messages and a restart prompt when work is running (#3928)
- The top bar is more compact: the account button is just your profile photo, or a person icon when signed out, and the GitHub star count is shortened (#3922)
- Product analytics moved from PostHog to Supermetry, still only after a yes at the first-run prompt, and now also records which engine ran each generation and transcription (#3902)
- Restyled Settings → VoiceStudio account to match the home page: one account card with your plan and storage, the email shown once, and a storage bar visible at 0% (#3885)
- Account sign-in is available in every build, still off until you turn it on, and now starts on the VoiceStudio dashboard with a separate Sign up option (#3850)
- The Pro page in the app now shows the correct $199 Lifetime price (#3831)
- Mac releases use an Apple API key for notarization instead of an account password (#3775) — thanks @velixio!
- Source development now prepares the local API environment before launching Electron (#3637)
- Remove obsolete screenshot and source-archive helpers that targeted the retired UI layout (#3623)
- Removed unused development tools (pytest-cov, Turborepo and a duplicate TypeScript) from the repository and fixed the remaining broken documentation links (#3571, #2603, #2903) — thanks @themrbeasley and @marciok!
- Workflows open at /workflows framed to their steps, Calls opens at /calls, and an unknown address shows Page not found (#3547)
- Projects lists projects under All projects with voices, transcripts, takes and exports grouped under Library; saving a dub now lives in Dubbing (#3547)
- The OmniVoice sidecar availability check no longer imports the model, so backend startup reconcile finishes about 6 s sooner (#3534) — thanks @adrianchen91!
- The first-run install screen is cleaner: one progress bar, a step checklist that shows the package being installed, and Cancel and the log inside the progress card (#3534) — thanks @meisolated!
Added
v0.5.8-beta.5 — VoiceStudio预发布2026-11-12
Installer signing
| Platform | Signing |
|---|---|
| macOS (Apple Silicon) | ✅ Signed with Developer ID and notarized by Apple |
| macOS (Intel) | ✅ Signed with Developer ID and notarized by Apple |
| Windows x64 | ⚠️ Unsigned: SmartScreen shows "Windows protected your PC"; choose More info → Run anyway |
| Windows ARM64 | ⚠️ Unsigned: SmartScreen shows "Windows protected your PC"; choose More info → Run anyway |
| Linux x64 | — Linux packages are not code-signed |
Phone downloads
Android: signed APK · Play bundle.
iPhone: unsigned sideload IPA. Sign with your own Apple ID to install; this is not a TestFlight build.
Windows avoids the crash while loading models, Intel Arc GPUs run OmniVoice, and OmniVoice GGUF resident mode now works in release builds. OmniVoice GGUF runs on AMD, Intel and NVIDIA GPUs through Vulkan on Windows and Linux, dots.tts arrives on Windows with no Python setup, and a failing PyTorch download or a broken GPU no longer leaves you without a working app.
New since beta.4: Windows loads model weights one at a time by default and names the library behind a crash, and Restart on CPU no longer touches the GPU driver; installers bundle tts-server, so OmniVoice GGUF resident mode (Settings → Performance) works; Intel Arc GPUs run OmniVoice in bfloat16; CosyVoice 3 GGUF finds your Vulkan GPU again; the watermark no longer adds a faint offset to silences; Ascend NPUs are detected again; a new macOS installer window; and a default install about 81 MB smaller.
Testing wanted: these get their first real-hardware test in this beta, so reports are welcome: OmniVoice GGUF resident mode with the bundled tts-server, Intel Arc (XPU) GPUs, Restart on CPU after a graphics driver crash, the new Windows default of loading model weights one at a time, and the new macOS DMG window.
The new macOS installer window: drag VoiceStudio to Applications (preview built from the installer's own background and icons)
Thanks: @mrnitrox (Intel Arc support, #4022) and @khawajaJunaid (Intel Mac Radeon docs, #4014) for code; @tera-png (#3980), @Imammahmuda1804 (#3767), @oryx-system (#3772), @pimeNtex (#4024) and @Patrix9999 (#2818) for bug reports.
Download
| Platform | Installer |
| --- | --- |
| Windows x64 | Installer |
| Windows ARM64 (experimental) | Installer |
| macOS Apple Silicon | DMG |
| macOS Intel | DMG |
| Linux x64 | AppImage · deb |
| Android | APK · AAB |
| iPhone | Unsigned IPA (sideload with your own Apple ID) |
Upgrading: Install over your existing app; voices, projects and settings are kept.
Highlights
- OmniVoice GGUF uses AMD, Intel and NVIDIA GPUs through Vulkan on Windows and Linux
- dots.tts voice cloning now runs on Windows through audio.cpp, with no Python setup
- Convert can keep your delivery, changing only the voice
- Dubbing exports subtitles before dubbing and moves lines between speakers
- Setup falls back to the CPU build when PyTorch downloads fail, and the app falls back to the CPU when the GPU fails
Changed
- Smaller install: Gradio and WebDataset are no longer installed by default (about 81 MB less); the Gradio demo and model training now use the ui and train extras (#4027)
- The macOS installer window has a new branded background, larger icons and a clean volume name without the processor suffix (#4001)
- The backend starts lighter and faster: speech-recognition engines are checked in a short-lived helper process instead of being loaded at startup, while broken installs are still reported as unavailable (#3924, #3946)
- The app opens with less code to load: the audio and video players load with the first clip played and returning users skip the first-run setup screens, cutting the startup bundle by about 6%, and a new offline benchmark with a CI check on bundle growth tracks startup, memory, CPU and bundle size (#3924, #3940, #3943)
- Settings → VoiceStudio account lists what to sync and your devices as compact, aligned rows with icons, a Pro badge with one Get Pro link for locked items, and device rows that show OS and when each was last active (#3937)
- The runtime setup screen fits one window: one status line, the install folder with free space, one main action (or the exact fix, such as choosing another folder when the drive is full), with language, region, Python and remote options under Options and recovery tools under More fixes (#3934)
- OmniVoice GGUF can keep its runtime loaded between requests to skip the reload before each generation, as an opt-in under Settings → Performance (#3965, #3654)
- The audio.cpp runtime updates to v0.9.4; installed Breeze-TTS-2, dots.tts and Sortformer models keep working without a re-download, and the previous runtime keeps running them until you install the update (#3129)
- Updates now download in the background and install when you quit, with a compact title-bar pill for progress, Restart and Retry, clearer error messages and a restart prompt when work is running (#3928)
- The top bar is more compact: the account button is just your profile photo, or a person icon when signed out, and the GitHub star count is shortened (#3922).
- Product analytics moved from PostHog to Supermetry, still only after a yes at the first-run prompt, and now also records which engine ran each generation and transcription. (#3902)
- Restyled Settings → VoiceStudio account to match the home page: one account card with your plan and storage, the email shown once, and a storage bar visible at 0%. (#3885)
- Account sign-in is available in every build, still off until you turn it on, and now starts on the VoiceStudio dashboard with a separate Sign up option (#3850)
- The Pro page in the app now shows the correct $199 Lifetime price (#3831)
- Mac releases use an Apple API key for notarization instead of an account password. (#3775) — thanks @velixio!
- Source development now prepares the local API environment before launching Electron (#3637)
- Remove obsolete screenshot and source-archive helpers that targeted the retired UI layout. (#3623)
- Removed unused development tools (pytest-cov, Turborepo and a duplicate TypeScript) from the repository and fixed the remaining broken documentation links (#3571, #2603, #2903) — thanks @themrbeasley and @marciok!
- Workflows open at /workflows framed to their steps, Calls opens at /calls, and an unknown address shows Page not found (#3547)
- Projects lists projects under All projects with voices, transcripts, takes and exports grouped unde
v0.5.8-beta.4 — VoiceStudio预发布2026-11-10
Installer signing
| Platform | Signing |
|---|---|
| macOS (Apple Silicon) | ✅ Signed with Developer ID and notarized by Apple |
| macOS (Intel) | ✅ Signed with Developer ID and notarized by Apple |
| Windows x64 | ⚠️ Unsigned: SmartScreen shows "Windows protected your PC"; choose More info → Run anyway |
| Windows ARM64 | ⚠️ Unsigned: SmartScreen shows "Windows protected your PC"; choose More info → Run anyway |
| Linux x64 | — Linux packages are not code-signed |
Phone downloads
Android: signed APK · Play bundle.
iPhone: unsigned sideload IPA. Sign with your own Apple ID to install; this is not a TestFlight build.
AMD and Intel GPUs now speed up OmniVoice, and setup gets far more forgiving. OmniVoice GGUF runs on AMD, Intel and NVIDIA GPUs through Vulkan on Windows and Linux, dots.tts arrives on Windows with no Python setup, and a failing PyTorch download or a broken GPU no longer leaves you without a working app.
New since beta.3: audio.cpp v0.9.4, hardware video encoders for dub exports, the Ming-3.0-tts and MOSS-SoundEffect engines, a voicestudio terminal command, streaming OpenAI-compatible speech, Windows ROCm for RDNA 1 and 2 GPUs, per-voice pronunciation dictionaries, Convert voice for video, source-language dubbing, a lighter startup, a tidier setup screen and account page, and dozens of dubbing, setup, Snap and Android fixes.
Testing wanted: these get their first real-hardware test in this beta, so reports are welcome: the audio.cpp v0.9.4 runtime, hardware video encoders for dub exports, the Ming-3.0-tts engine, sound effects (MOSS-SoundEffect, NVIDIA GPUs only), the Snap fixes and the Windows uninstall prompt that offers to delete downloaded models. OmniVoice GGUF resident mode is opt-in and inactive until a build ships the tts-server binary.
Download
| Platform | Installer |
| --- | --- |
| Windows x64 | Installer |
| Windows ARM64 (experimental) | Installer |
| macOS Apple Silicon | DMG |
| macOS Intel | DMG |
| Linux x64 | AppImage · deb |
| Android | APK · AAB |
| iPhone | Unsigned IPA (sideload with your own Apple ID) |
Upgrading: Install over your existing app; voices, projects and settings are kept.
Highlights
- OmniVoice GGUF uses AMD, Intel and NVIDIA GPUs through Vulkan on Windows and Linux
- dots.tts voice cloning now runs on Windows through audio.cpp, with no Python setup
- Convert can keep your delivery, changing only the voice
- Dubbing exports subtitles before dubbing and moves lines between speakers
- Setup falls back to the CPU build when PyTorch downloads fail, and the app falls back to the CPU when the GPU fails
Changed
- The backend starts lighter and faster: speech-recognition engines are checked in a short-lived helper process instead of being loaded at startup, while broken installs are still reported as unavailable (#3924, #3946)
- The app opens with less code to load: the audio and video players load with the first clip played and returning users skip the first-run setup screens, cutting the startup bundle by about 6%, and a new offline benchmark with a CI check on bundle growth tracks startup, memory, CPU and bundle size (#3924, #3940, #3943)
- Settings → VoiceStudio account lists what to sync and your devices as compact, aligned rows with icons, a Pro badge with one Get Pro link for locked items, and device rows that show OS and when each was last active (#3937)
- The runtime setup screen fits one window: one status line, the install folder with free space, one main action (or the exact fix, such as choosing another folder when the drive is full), with language, region, Python and remote options under Options and recovery tools under More fixes (#3934)
- OmniVoice GGUF can keep its runtime loaded between requests to skip the reload before each generation, as an opt-in under Settings → Performance; it stays inactive until a build ships omnivoice.cpp's tts-server, so nothing changes today (#3965, #3654)
- The audio.cpp runtime updates to v0.9.4; installed Breeze-TTS-2, dots.tts and Sortformer models keep working without a re-download, and the previous runtime keeps running them until you install the update (#3129)
- Updates now download in the background and install when you quit, with a compact title-bar pill for progress, Restart and Retry, clearer error messages and a restart prompt when work is running (#3928)
- The top bar is more compact: the account button is just your profile photo, or a person icon when signed out, and the GitHub star count is shortened (#3922).
- Product analytics moved from PostHog to Supermetry, still only after a yes at the first-run prompt, and now also records which engine ran each generation and transcription. (#3902)
- Restyled Settings → VoiceStudio account to match the home page: one account card with your plan and storage, the email shown once, and a storage bar visible at 0%. (#3885)
- Account sign-in is available in every build, still off until you turn it on, and now starts on the VoiceStudio dashboard with a separate Sign up option (#3850)
- The Pro page in the app now shows the correct $199 Lifetime price (#3831)
- Mac releases use an Apple API key for notarization instead of an account password. (#3775) — thanks @velixio!
- Source development now prepares the local API environment before launching Electron (#3637)
- Remove obsolete screenshot and source-archive helpers that targeted the retired UI layout. (#3623)
- Removed unused development tools (pytest-cov, Turborepo and a duplicate TypeScript) from the repository and fixed the remaining broken documentation links (#3571, #2603, #2903) — thanks @themrbeasley and @marciok!
- Workflows open at /workflows framed to their steps, Calls opens at /calls, and an unknown address shows Page not found (#3547)
- Projects lists projects under All projects with voices, transcripts, takes and exports grouped under Library; saving a dub now lives in Dubbing (#3547)
- The OmniVoice sidecar availability check no longer imports the model, so backend startup reconcile finishes about 6 s sooner (#3534) — thanks @adrianchen91!
- The first-run install screen is cleaner: one progress bar, a step checklist that shows the package being installed, and Cancel and the log inside the progress card (#3534) — thanks @meisolated!
Added
- Qwen Base can deliver audio during generation on the speech WebSocket (#3992) — thanks @kellerkind84!
- OpenAI-compatible PCM speech can stream during synthesis, with native Qwen3-TTS Base support on Apple Silicon and provisioned NVIDIA systems (#3976) — thanks @kellerkind84!
- The account menu in th
CTranslate2 4.8.4 HIP (Windows ROCm 10, RDNA 2)预发布2026-11-09
Built by .github/workflows/ct2-rocm-windows.yml from OpenNMT/CTranslate2 78b1e1770fd3c28dbd125bbe586d90ae45e611b8 + cf4579017d089f49446edc59639f58d88ea79880 (tree 7d3ab96d381215943951607689fce10b32e192d8). Dependency asset for the opt-in Windows ROCm setup; not a VoiceStudio release.