Your voice. Your engines. Running where your data lives.
RealVoice is the voice modality of the Hominis multi-modal stack. Clone a voice, synthesize speech, dictate anywhere and give your agents a voice · local-first, on infrastructure you own.
Powered by Hominis AIYour clone · Qwen3-TTS · EN
“This is my voice. I recorded it once, and RealVoice does the rest.”
Your clone voice sample paused.
The voice in the Hominis multi-modal stack.
Hominis is a multi-modal foundation · RealVoice is its voice. The same stack that gives agents procedural competence now speaks and listens, local-first. Clone once, use it across every RealAI surface.
One voice layer, end to end.
Voice cloning
Clone a voice from a short reference sample · zero-shot, in a voice you own and control.
Neural synthesis
Six production engines behind one interface. Pick the right one per job.
Dictation
A global hotkey turns speech into text in whatever app has focus.
Agent voice
Any MCP-aware agent speaks in a voice you bound to it.
Real-time streaming
Audio streams as it generates · low latency for live apps.
Own-your-engine deploy
Desktop, private VPC or air-gapped. No black box.
Press play. Generated locally, in a voice we own.
Each clip is produced on one machine, from a single recording. More voices arrive as we build their presets.
Your clone
Qwen3-TTS · EN
“This is my voice. I recorded it once, and RealVoice does the rest.”
Narrator
Qwen CustomVoice · EN
Voice sample coming soon.
Support agent
Qwen CustomVoice · EN
Voice sample coming soon.
EU voice
Chatterbox ML · NL
Voice sample coming soon.
Three steps from sample to voice.
Record
Capture 20–40 seconds of clean reference audio · your voice, or a licensed talent.
Model
RealVoice builds a voice profile locally · no audio leaves your machine.
Voice
Synthesize unlimited speech in that voice, in real time, across every RealAI surface.
Any clean take becomes a voice.
Upload, record or capture. The profile is built on your machine either way.
Upload a file
Drop in an existing recording. A single clean take is enough to build a profile.
Record live
Speak into your microphone and clone from the take, without leaving the studio.
Capture system audio
Pull a clip from whatever your machine is playing, with consent and rights in hand.
Speak into any application.
A global hotkey, local transcription, and the text lands where your cursor already was.
Press
A global hotkey starts capture, push-to-talk or toggle. It works over any application.
Transcribe
Whisper runs on your machine and turns the take into text.
Refine
An optional local model cleans up filler, punctuation and technical terms.
Paste
The text lands in whatever field had focus. Nothing was uploaded.
Give your agents a voice you own.
One endpoint, and the agent speaks
RealVoice exposes an MCP endpoint. Bind a voice to an agent once, and every tool call it makes can be spoken back in that voice. The same endpoint transcribes audio, so an agent can listen as well as speak. It all resolves on your machine.
A sellable set of production voice engines.
Pick the engine that fits the job. Every one in the paid build is licensed for resale.
| Engine | Voice | Languages | Best for |
|---|---|---|---|
| Qwen3-TTS Base | Zero-shot clone | 10 | Delivery instructions in natural language |
| Qwen CustomVoice | Preset voices | 10 | Style control without a reference sample |
| Chatterbox Multilingual | Zero-shot clone | 23 | The broadest language coverage |
| Chatterbox Turbo | Zero-shot clone | English | Fast, with paralinguistic tags |
| LuxTTS | Zero-shot clone | English | Light footprint, 48 kHz output |
| TADA | Zero-shot clone | 10 | Expressive flow-matching synthesis |
Kokoro is not in the paid build. Its phonemizer chain carries GPL and LGPL obligations that do not suit a product you resell. TADA ships under the Llama 3.2 community licence and carries the Built with Llama attribution.
Twenty-three languages, one engine.
The multilingual engine covers most of Europe and reaches well beyond it.
The other engines cover ten languages, or English at higher speed.
A local endpoint your code can call.
RealVoice runs a service on your machine. Everything the studio does is available over HTTP, and to any MCP-aware agent.
POST /generate/stream
Synthesize speech and receive the audio in one request.
POST /generate
Queue a long generation and poll it to completion.
POST /profiles
Create a voice profile.
POST /profiles/{id}/samples
Attach a reference sample and clone the voice.
POST /transcribe
Turn recorded audio into text.
POST /speak
Speak in the voice bound to the calling agent.
Synthesize in a cloned voice
curl -X POST http://127.0.0.1:17493/generate/stream \
-H "content-type: application/json" \
-d '{"profile_id":"<your-voice>","text":"Deploy complete.","engine":"qwen"}' \
--output speech.wavThe same service answers on an MCP endpoint, so an agent can speak, transcribe and list the voices it is allowed to use.
Run it where your data already lives.
Apple Silicon, NVIDIA, AMD and Intel GPUs, with a CPU fallback.
On the desktop
A native app that runs the engines on your own machine. Nothing to provision.
In a private VPC
Run the service inside your perimeter and call it over your own network.
Air-gapped
Download the models once, then run with no outbound connection at all.
Your voice never leaves your infrastructure.
Runs local
Inference happens on your machine or in your VPC. The only outbound call is the one-time model download.
Own your voice
Voice profiles and samples are stored on disk, never uploaded, never used to train someone else's model.
EU-resident
Deploy in EU regions for GDPR by default, or run with no network at all.
Sellable engines
Every engine in the paid build is MIT or Apache-2.0 licensed, so you can ship it inside your own product.
Questions about RealVoice
RealVoice is the voice modality of the Hominis multi-modal stack · the same foundation that gives agents procedural competence, now able to speak and listen. It runs local-first alongside the model and the Agentic OS.
No. RealVoice runs inference on your machine or your own infrastructure by default. Voice samples, generated audio and profiles stay local. The only outbound request is the one-time model download.
A single clean take of roughly twenty to forty seconds, plus its transcript. The profile is built locally and can be used immediately.
Six production engines · Qwen3-TTS Base, Qwen CustomVoice, Chatterbox Multilingual, Chatterbox Turbo, LuxTTS and TADA.
Kokoro's grapheme-to-phoneme chain depends on GPL and LGPL components, which do not suit a product our customers resell. We leave it out of the paid build rather than pass that obligation on.
Yes. Every engine in the paid build is permissively licensed, so you can ship RealVoice and the voices you create inside your own commercial product. TADA ships under the Llama 3.2 community licence and carries its attribution.
Twenty-three languages through the multilingual engine, including most of Europe plus Arabic, Hindi, Japanese, Korean, Chinese and Swahili. The other engines cover ten languages or English.
Yes. RealVoice exposes an MCP endpoint, so any MCP-aware agent can speak in a voice you bound to it, and transcribe audio, without leaving your machine.
On the desktop, inside a private VPC, or fully air-gapped. Apple Silicon, NVIDIA, AMD and Intel GPUs are supported, with a CPU fallback.