FIG. 01
PLUGGABLE ENGINES
Seven open-source engines — VibeVoice, Kokoro, Kitten TTS Mini, Chatterbox, OmniVoice, VoxCPM, and Qwen3-TTS — behind one clean interface. Only one loads at a time to keep memory low.
YOUR SCRIPTLOCAL AUDIO
ENGINE: LOADED//AUDIO STAYS LOCAL
Self-Hosted,OpenSource
alternativetoElevenLabs
Seven open-source TTS engines — voice cloning, multi-speaker podcasts, and 30+ languages. All on your machine. No cloud, no API keys, and audio never leaves your computer.
7 TTS ENGINES30+ LANGUAGESFULLY OFFLINE
// SEVEN OPEN-SOURCE TTS ENGINES · ONE LOCAL STUDIO
VibeVoice
Kokoro
Kitten TTS
Chatterbox
OmniVoice
VoxCPM2
Qwen3-TTS
FastAPI
PyTorch
Hugging Face
Voice Cloning
100% Offline
30+ Languages
VibeVoice
Kokoro
Kitten TTS
Chatterbox
OmniVoice
VoxCPM2
Qwen3-TTS
FastAPI
PyTorch
Hugging Face
Voice Cloning
100% Offline
30+ Languages
// CORE_CAPABILITIES
EVERYTHING, ON YOUR MACHINE
A full text-to-speech studio that runs locally — multi-engine, multi-speaker, and private by design.
FIG. 02
FULLY OFFLINE
No cloud, no telemetry, no API keys. Audio never leaves your machine.
FIG. 03
VOICE CLONING
Clone a voice from a short reference clip, or design one from a text prompt.
FIG. 04
VOICE MODES
Clone, design, or auto — per speaker.
FIG. 05
30+ LANGUAGES
Multilingual synthesis with native-script support: اردو, हिन्दी, 中文 and more.
FIG. 06
IN-APP INSTALLS
Install isolated engines and download model weights from the UI, with live progress.
FIG. 07
AUTO-UPDATES
Checks GitHub releases and applies updates in place, right from the app.
FIG. 08
FOUR PROJECT MODES
Podcast, Text-to-Voice, Transcribe (speech-to-text with Whisper), and Dub — re-voice any clip in a voice you choose.
FIG. 09
VOICE LIBRARY
Manage built-in and cloned voices with per-voice metadata and reference transcripts.
// HOW_IT_WORKS
FROM SCRIPT TO SPEECH
A pluggable engine pipeline: pick voices, load one engine, synthesize locally, and keep every byte of audio on your machine.
● 100% LOCALNO CLOUD
01_INPUT
SCRIPT & VOICES
Write a multi-speaker script or plain text, then assign a voice per speaker — built-in or cloned from a reference clip.
PODCASTMULTI
TEXT-TO-VOICESINGLE
VOICESCLONE / BUILT-IN
02_ENGINE [CORE]
PLUGGABLE ENGINE
The selected engine loads lazily on first use — only one at a time to keep memory low. Conflicting engines run in their own isolated venv.
ACTIVE1 ENGINE
ISOLATIONPER-VENV
LOADLAZY
03_SYNTHESIZE
LOCAL SYNTHESIS
Multi-speaker lines render as separate calls and join with silence gaps. Results are cached to disk so repeats are instant.
SPLITPER-LINE
CACHEON DISK
GPU/CPUAUTO
04_PRIVATE
STAYS ON DEVICE
Audio is written straight to your disk. Nothing is uploaded, streamed to a cloud, or tracked — ever.
NETWORKNONE
TELEMETRYOFF
API KEYS0
// TTS_ENGINES
SEVEN ENGINES, ONE STUDIO
Each engine has its own strengths. Only one loads at a time to keep memory low — install and download any of them right from the app.
MULTI-SPEAKER
VibeVoice-1.5B
Expressive multi-speaker synthesis for rich, natural dialogue.
~5.4 GBEnglish
LIGHTWEIGHT
Kokoro-82M
Fast and tiny, yet multilingual — great on modest hardware.
~350 MB4 languages
RUNS ANYWHERE
Kitten TTS Mini
An ~80M ONNX model that runs on CPU with no GPU — the low-end-hardware tier.
~79 MBEnglish
CLONING
Chatterbox V3
Voice cloning across a wide range of languages.
~500 MB23 languages
VOICE MODES
OmniVoice
Clone from a clip, design from a prompt, or go fully automatic.
~3.3 GBMultilingual
HI-FI
VoxCPM2
High-fidelity 48 kHz output with ultimate reference cloning.
~5 GB30 languages
BUILT-IN VOICES
Qwen3-TTS
Nine premium built-in voices with free-text style control.
~1.7B params10 languages
RUNS LOCALLYSWITCH ANYTIMEONE ACTIVE ENGINE
// UNDER_THE_HOOD
BUILT ON OPEN SOURCE
A FastAPI backend serving open-source models to a React editor — all running on your own hardware, no accounts required.
FastAPI0.115+
Backend API server
React + Vite18
Frontend editor UI
PyTorch2.4+
Model inference
Hugging Face
Model weights & cache
Transformers4.51
Engine runtimes
CUDA · MPS · CPU
Auto-detected acceleration
SYNTHESIS ENGINE
RUNNING LOCALENGINE
1 ACTIVE
DEVICE
AUTO
CACHE
ON DISK
// LOG_STREAM
LIVE00:01> load_engine(name="vibevoice")
00:01> device=cuda --dtype=fp16DONE
00:02> tokenize --speakers=3OK
00:02> synthesize --segment=1/4OK
00:03> cache_write --format=wavOK
00:03> join --silence-gap=250msOK
Local-only
No network calls
Uvicorn0.30+
ASGI server
Pydantic2.7+
Config & schemas
Python3.10+
Backend language
Tailwind CSS3.4
App styling
Disk cache
Synthesis cache
// BY_THE_NUMBERS
AT A GLANCE
LOCAL · PRIVATESTATUS: OFFLINE
0
TTS ENGINES
all open source
0+
LANGUAGES
multilingual synthesis
0kHz
SAMPLE RATE
VoxCPM2 hi-fi
0%
OFFLINE
no cloud, no API keys
// COMPLIANCE_LOGFULLY OFFLINE[VERIFIED]OPEN SOURCE[ACTIVE]NO API KEYS[COMPLIANT]
// GET_STARTED
UP AND RUNNING IN TWO COMMANDS
Python 3.10+ and Node 18+. Setup auto-detects your GPU and installs the matching PyTorch build, then lets you pick which models to download.
git clone https://github.com/msrbuilds/voice-studio.git
cd voice-studio
python studio.py setup # venv, deps, CUDA auto-detect, model picker
python studio.py start # backend + frontend1Clone the repo2python studio.py setup3python studio.py start
100% OFFLINE • NO API KEYS • OPEN SOURCE