VoiceStudio

VoiceStudio is a desktop voice creation app open-sourced by Palash Debnath on GitHub and maintained by the company Yupcha Softwares, with versions for macOS, Windows and Linux. Open Voice cloning, pick a voice or add a clean reference recording, type your text and generate. You can also describe a voice in words (voice design), dub a video with timed, translated speech, or turn long text into an audiobook and queue batch jobs. It also has a floating dictation widget, transcription, vocal isolation and speaker diarization. The default speech model is k2-fsa's OmniVoice, and you can switch to more than a dozen other engines such as CosyVoice and IndexTTS. The README's headline says 646 languages, but which languages actually work depends on the engine and model you pick. The app serves an OpenAI-compatible API and an MCP server on local port 3900, and the repo's voicestudio Skill lets agents such as Claude Code and Codex call it to generate and transcribe audio. What sets it apart is how openly it lays out model licensing: the Model Manager lists each model's license source, evidence links and unresolved gaps, and separates commercial use of the model, commercial use of the generated audio, redistribution, and voice and recording consent.

Who it's for:creators who want to make narration, video dubs, audiobooks or dictation on their own computer, and who are willing to manage model downloads, hardware needs and model licenses themselves.

Before you start: get the person's consent before you use someone else's voice. The README says: "Clone voices only with permission." The app lets you record a consent statement for a voice profile and revoke it later; the source code describes this recording as a provenance record, not identity or voiceprint verification, and ordinary local generation doesn't require it. Read the licenses separately. The app itself is AGPL-3.0 and can be used commercially under AGPL terms; embedding it in closed-source products needs a paid commercial license. The default OmniVoice model's weights are CC-BY-NC, for non-commercial use only, and its audio tokenizer carries separate Higgs Audio and Llama community terms. The project's own model license inventory covers 56 models and clears none of them for commercial use: some are known non-commercial and the rest are still unreviewed, a licensing gap the project itself acknowledges. A paid app license doesn't cover rights to the models, the generated audio or the voices. You can install from the macOS, Windows and Linux packages or run the one-line install script from the official site, which checks SHA-256. It works on the CPU without a dedicated GPU; the README says the CPU build needs about 5 GB of space, and model downloads can run to several GB. Intel Macs get the interface only and must connect to a remote backend. Recording and dictation need microphone permission, and the system permissions needed to insert dictated text into other apps vary by platform; we haven't verified them. The README says local workflows run on your own hardware and remote services are optional: online translation, remote models or backends and outside providers all need your own setup, may involve accounts, keys and fees, and send your content to those services. By default the backend accepts only local connections and has no login. Turning on network sharing creates a 6-digit PIN, and remote access can also use an API key; we suggest not exposing it directly to the internet. Usage analytics are off by default, and only after you agree are fixed screen names and action labels sent to PostHog. Generation history has a retention cap, and older unstarred takes and their audio are deleted once it's reached.

Based on the project's public documentation. Not tested by AI Inspiration World.