Description
Narrating a video through an online TTS service bills you per character, so five passes over a two-thousand-word script is five invoices — and getting a voice that fits means uploading your reference audio to someone else's servers. yovoice moves the whole thing onto your machine: the model runs on your own GPU, with no cloud API and no per-character charge, a reference clip under a minute is enough to clone a voice, and the generated speech can be directed for emotion.
It is a full workspace rather than a wrapper around a script — projects on the left, the script in the middle, voice, delivery mode and emotion strength on the right, and a waveform with a player along the bottom, so rewriting one line means regenerating one line while earlier versions stay in history. Delivery can follow a reference voice, match a reference performance, use emotion presets, or simply be described in words. Packages for macOS 14+ on Apple Silicon and Windows 10/11, Apache-2.0 licensed.
Local generation, no metering: models and inference stay on your machine with no cloud API in the loop, so revising a script as many times as you like costs nothing extra and the text never leaves the computer.
Voice cloning: import or record a 1-to-60-second reference clip to clone a voice, while VoxCPM2 also offers text-guided voice design with no reference at all, plus transcript-assisted controllable cloning.
Emotion and delivery control: four expression modes — follow the reference voice, match a reference performance, adjust emotion presets, or describe the delivery in words — with a continuous slider for emotional intensity rather than one flat reading voice.
Several local models: IndexTTS 2.0 / 2.5 and VoxCPM2 run through audio.cpp, models download inside the app with resumable transfers, and your own GGUF weights can be imported.
Hardware acceleration: Metal on Apple Silicon, and CPU, NVIDIA CUDA or experimental Vulkan on Windows, with the CUDA engine downloadable from Settings.
A complete audio workflow: import or record reference audio, trim clips, preview speech and export the result, with work organized into projects and a history of earlier generations to compare against.
48 kHz output: VoxCPM2 renders at 48 kHz and handles multilingual text automatically, so the result does not announce itself as low-rate machine speech in a video.
CLI and Agent Skill: a standalone command-line tool runs without opening the desktop app, and the bundled Agent Skill lets an AI assistant install the engine and model itself, then produce a voiceover from a script and a reference clip on a single instruction.
Data stays local: projects, voices and settings live in a .yovoice folder in your home directory, survive uninstalling, and move to another machine by copying that folder.
It is a full workspace rather than a wrapper around a script — projects on the left, the script in the middle, voice, delivery mode and emotion strength on the right, and a waveform with a player along the bottom, so rewriting one line means regenerating one line while earlier versions stay in history. Delivery can follow a reference voice, match a reference performance, use emotion presets, or simply be described in words. Packages for macOS 14+ on Apple Silicon and Windows 10/11, Apache-2.0 licensed.
Features
Local generation, no metering: models and inference stay on your machine with no cloud API in the loop, so revising a script as many times as you like costs nothing extra and the text never leaves the computer.
Voice cloning: import or record a 1-to-60-second reference clip to clone a voice, while VoxCPM2 also offers text-guided voice design with no reference at all, plus transcript-assisted controllable cloning.
Emotion and delivery control: four expression modes — follow the reference voice, match a reference performance, adjust emotion presets, or describe the delivery in words — with a continuous slider for emotional intensity rather than one flat reading voice.
Several local models: IndexTTS 2.0 / 2.5 and VoxCPM2 run through audio.cpp, models download inside the app with resumable transfers, and your own GGUF weights can be imported.
Hardware acceleration: Metal on Apple Silicon, and CPU, NVIDIA CUDA or experimental Vulkan on Windows, with the CUDA engine downloadable from Settings.
A complete audio workflow: import or record reference audio, trim clips, preview speech and export the result, with work organized into projects and a history of earlier generations to compare against.
48 kHz output: VoxCPM2 renders at 48 kHz and handles multilingual text automatically, so the result does not announce itself as low-rate machine speech in a video.
CLI and Agent Skill: a standalone command-line tool runs without opening the desktop app, and the bundled Agent Skill lets an AI assistant install the engine and model itself, then produce a voiceover from a script and a reference clip on a single instruction.
Data stays local: projects, voices and settings live in a .yovoice folder in your home directory, survive uninstalling, and move to another machine by copying that folder.
