Seven speech engines share one local studio
Voicebox combines voice cloning, text-to-speech, transcription, global dictation, effects, and a multitrack story editor. Its 7 documented speech engines cover different languages and controls, while Whisper handles speech-to-text. Generated takes retain their source history, profiles can hold several reference samples, and longer scripts are split and crossfaded.
The breadth produces a large codebase. Our checkout had 681 files, about 72,798 lines of source, and occupied 54.9 MB before installation. React and TypeScript power the interface, Tauri and Rust handle the desktop layer, and a Python FastAPI server runs the speech stack. Bun workspaces cover the app, web version, landing site, and Tauri package.
REST and MCP turn the desktop app into local voice plumbing
The local server exposes generation, speaking, transcription, and profile routes on port 17493. Its built-in MCP server adds voicebox.speak, voicebox.transcribe, voicebox.list_captures, and voicebox.list_profiles. An HTTP client ID can bind a different default voice to each agent.
The 911 MB occupied by our 408 installed packages arrives before the speech, transcription, and local language-model downloads. A user choosing among MLX, CUDA, ROCm, DirectML, XPU, and CPU must match the backend to the machine and check that the app selected the expected path.
What happened when we ran it
Our sandbox installed 408 packages with Bun in 50 seconds and used 911 MB on disk. We tested commit 51f49de in a fresh unprivileged Debian container with 3 CPUs, 8 GB of RAM, Node 22, and no secrets. The dependency step succeeded.
The build failed after 10 seconds with exit code 127. Its last lines show build-server.sh announcing an x86_64 Linux sidecar, attempting to install PyInstaller, and then reporting python: command not found. The package build exited with code 127. No tests script or target existed, so the lab skipped tests.
Python 3.11 and Tauri packages are part of the source build
The development guide requires Bun, Python 3.11 or newer, Rust, Tauri's operating-system prerequisites, Git, and the just command runner. Xcode joins the list on macOS. Production builds turn the Python server into a sidecar before Tauri packages the native application.
Our build reached that sidecar step only 10 seconds after the 50-second install, then stopped at the missing python command. The README does list Python clearly, so this is a prerequisite mismatch in our sandbox rather than a hidden requirement.
Cached models can work offline while the desktop UI waits for Wi-Fi
Voicebox describes its models, captures, and voice data as local. Issue 1085 narrows an awkward exception: with a model fully cached on Apple Silicon, the desktop Generate button reportedly spins until Wi-Fi returns. The reporter says the backend continues answering local polling and a direct curl request generates audio offline.
The lab did not launch the application or download models; its 911 MB result covers the 408 installed packages only. Before relying on offline use, disconnect the target machine and test generation, transcription, profile loading, and export through the same interface users will touch.
Non-English output has current correctness reports
Issue 1086 reports German output turning into noise or nonsense with Qwen TTS 1.7B for several voices. Issue 1066 documents a Hindi script being cut off while the generation still reported completion, linking the behavior to a character-based chunk limit. Issue 603 reports Korean dictation refinement returning English instead of cleaned Korean.
With 72,798 source lines and 7 speech engines, a single green status cannot certify the spoken result. Teams should compare generated audio with the full input, check language and duration, and listen for noise before publishing. This is especially important for long scripts and unattended API jobs. Voicebox keeps raw captures and generation versions, which helps investigation, but provenance cannot repair audio that was truncated or spoken in the wrong language.
Linux requires source work, and auto-paste is still platform-specific
Voicebox publishes desktop downloads for macOS and Windows plus a Docker route. The README says Linux desktop binaries are not yet available. Global dictation can record across platforms, but automatic paste into the previously focused field is documented for macOS. Windows and Linux auto-paste appears in the roadmap, so it should be treated as an aspiration until shipped rather than a current feature.
The repository includes 3 CI workflow files, a Dockerfile, and a Compose file, but no tests directory. Our lab found no test target and could not complete the build at commit 51f49de. Docker may suit API use, while Tauri is needed for native shortcuts, overlays, and desktop permissions. Decide which half you need before installing: a headless speech service and a global dictation app have different platform requirements and failure modes.
August code activity coexists with a large support queue
GitHub recorded the last push on August 9, 2026, and release v0.5.0 was published on April 25. The repository had 52,766 stars and 680 combined open issues and pull requests when fetched. Issues were still receiving updates in September, including fresh reports about audio correctness and offline use, so the later issue activity matters alongside the earlier push date. The combined open count is not a count of confirmed bugs.
The 408-package install and failed 10-second build show both sides of Voicebox. There is serious product scope here, plus source and platform work that a polished interface can hide. Developers should trial one engine, one language, and one delivery path before adopting the whole studio. For API automation, verify the response body and the audio itself; for dictation, test permissions and offline behavior on the exact operating system.