Voicebox: A Local-First Voice Studio That Covers Both Input and Output

What Voicebox Solves and Who It Is For

Voicebox targets a gap in the voice AI market. ElevenLabs handles speech output, WisprFlow handles input, but neither does both, and both are cloud-based. Voicebox is a local-first alternative that combines voice cloning, speech generation, dictation, and agent voice output in one app. It is built for developers who want to integrate voice into their own applications via a REST API and MCP server, and for privacy-conscious users who do not want their voice data or captures leaving their machine. The README is explicit: models, voice data, and captures never leave your machine. This makes it suitable for workflows where confidentiality is a concern, such as medical dictation or legal transcription, though the project does not claim any compliance certification.

The Architecture: Tauri, Local Models, and a Bundled LLM

Voicebox is built with Tauri, which means the frontend uses web technologies but the backend is Rust, not Electron. This choice affects memory footprint and startup time, though the README does not provide benchmarks. The core loop is: you record or provide a reference audio sample, Voicebox clones the voice using one of seven TTS engines, and then you can generate speech. For dictation, Whisper-based speech-to-text runs locally. The bridge between input and output is a bundled local LLM that handles refinement and per-profile personas. This LLM enables modes like Compose, Rewrite, and Respond, which can also be invoked by agents over MCP. The architecture is modular: engines are switchable per-generation, and post-processing effects run through Spotify's pedalboard library. This modularity is a real strength, but it also means the user must understand which engine to choose for a given task.

Seven TTS Engines: How to Choose

Voicebox does not rely on a single TTS model. It offers seven engines, each with different language coverage and strengths. Qwen3-TTS is the high-quality multilingual option with 10 languages and supports delivery instructions like "speak slowly" or "whisper". Chatterbox Multilingual has the broadest coverage at 23 languages, including Arabic, Swahili, and Hindi. Chatterbox Turbo is a fast 350M model that is the only one that interprets paralinguistic tags like [laugh] and [sigh]. Kokoro is a tiny 82M model with 50 curated preset voices and fast CPU inference. LuxTTS is lightweight, requiring only about 1GB VRAM, and runs 150x realtime on CPU. TADA from HumeAI is a speech-language model that can produce 700 seconds or more of coherent audio. This variety is useful, but it also creates complexity. The user must know that if they type [laugh] with Qwen3-TTS, it will be read literally as text, not as an emotional cue. The README is clear about this limitation, but it is a trap for newcomers.

Getting Started: Commands and Configuration

Installation is straightforward on macOS and Windows. The README provides direct download links for DMG and MSI files. For Docker, the command is simply `docker compose up`. Linux users do not get prebuilt binaries; they must build from source using instructions on voicebox.sh/linux-install. This is a significant barrier for Linux users, especially since the project claims to support Linux, AMD ROCm, and Intel Arc. After installation, the first step is to download models, which the troubleshooting guide covers. Configuration is mostly through the GUI, but the API-first design means you can interact programmatically. The MCP server exposes a `voicebox.speak` tool call, which any MCP-aware agent can use. This is a concrete integration point: if you use Claude Code, Cursor, or Cline, you can give them a voice you have cloned. The repo does not show example API calls in the README, so you will need to consult the docs at docs.voicebox.sh for exact endpoints.

The Dictation and Agent Voice Output Workflow

Voicebox covers both sides of the voice loop. For input, it offers a global dictation hotkey with push-to-talk and toggle modes. On macOS, auto-paste is accessibility-verified, which is a practical detail for users with motor impairments. There is also an in-app mic on every text field. For output, the MCP server is the key feature. An agent can call `voicebox.speak` and the user hears the response in a chosen voice, including a cloned voice. This is different from typical TTS integrations because the voice can be a personal clone, not just a preset. The bundled LLM adds persona management: you can attach a free-form persona to a voice profile, and agents can invoke Compose, Rewrite, or Respond modes over MCP. This turns Voicebox into a voice interface for agents, not just a text-to-speech tool. The trade-off is that all of this runs locally, so you need a machine capable of running a local LLM alongside the TTS engines.

Limitations and Failure Modes

The most obvious limitation is the paralinguistic tag support. Only Chatterbox Turbo interprets tags like [laugh] and [sigh]. If you select any other engine, those tags are read literally. This is a documentation detail that can ruin a generation if you are not careful. Another limitation is Linux support: there are no prebuilt binaries, so Linux users must compile from source, which can be time-consuming and error-prone. The README also does not specify VRAM requirements for the larger models like Qwen3-TTS 1.7B or TADA 3B, so users with older GPUs may be unable to run them. The project is under active development, with releases every few days, which means the API and MCP interface may change. The troubleshooting guide exists, but the README does not list common issues, so you will need to read it for GPU and model-download problems. Finally, the promise of "complete privacy" is strong, but it relies on the user verifying that no telemetry is collected. The README states it, but an audit would be needed to confirm.

Alternatives and the Difference in Approach

The two named alternatives are ElevenLabs and WisprFlow. ElevenLabs is a cloud-based TTS service with high-quality voices and a large ecosystem, but it is not local and does not handle dictation. WisprFlow is a cloud dictation tool that focuses on voice input and transcription, but it does not offer voice cloning or TTS output. Voicebox combines both, but its local-first approach is the key difference. This means no data leaves your machine, but it also means you need to manage model downloads and GPU resources yourself. Another alternative is to use a dedicated local TTS like Piper or Coqui TTS, but those do not have the integrated dictation or MCP server. The MCP integration is a differentiator: it gives agents a voice without requiring a separate API key or network call. If you already use MCP-aware tools, Voicebox's built-in server is a direct integration point. If you do not need dictation or agent output, a simpler TTS tool might be easier to manage.

Maintenance and Upgrade Considerations

Voicebox is under active development, with version 0.5.0 released on 2026-04-25, following 0.4.5 and 0.4.4 within days. This pace suggests a responsive maintainer, but it also means upgrades are frequent. The README does not describe a migration path between versions, so you should check the changelog in the releases page. The license is MIT, which is permissive: you can use, modify, and distribute the code, including for commercial purposes, as long as you preserve the copyright notice. This is an advantage over some open-source projects that use AGPL. For maintenance, the main cost is model downloads and storage. Each TTS engine likely requires a separate model file, and the larger ones (Qwen3-TTS 1.7B, TADA 3B) will take up significant disk space and VRAM. The Docker option simplifies deployment, but it still requires a GPU passthrough for acceptable performance. Before adopting, verify that your hardware meets the requirements for the specific engines you plan to use, and budget time for regular updates.

Editorial conclusion

Voicebox is for developers and privacy-focused users who want a single local tool for both voice input and output, especially those already using MCP-aware agents like Claude Code or Cursor. It is not for users who need a fully polished commercial product with prebuilt Linux binaries or who rely on paralinguistic tags beyond Chatterbox Turbo. Before adopting, verify the GPU and VRAM requirements for your chosen engines, confirm the Linux build-from-source path if you are not on macOS or Windows, and test that the global dictation hotkey works reliably with your accessibility tools. The project is under active development with frequent releases, so check the latest changelog for breaking changes. If you need a specific language not covered by the engines you plan to use, confirm coverage first, since Chatterbox Multilingual is the only engine with 23 languages and others are limited to 8 to 10.

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes