Voila
Maitrix.org (with UC San Diego and MBZUAI) · Voice & Language
Voila is an open-source family of voice-language models from Maitrix.org that runs real-time, full-duplex voice conversations at 195 ms latency, with over a million built-in voices.
About Voila
What Is Voila
Voila is a research project, not a subscription app. A team at Maitrix.org, working with UC San Diego and MBZUAI, built a family of voice-language foundation models and released the code and weights for anyone to download. The goal is different from the voice assistants you already use. Instead of waiting for a command and answering it, Voila is built to listen, reason and speak at the same time.
Most voice assistants run on a chain. Speech recognition turns your words into text, a language model writes a reply, and a text-to-speech engine reads it back. Each link adds delay and strips out tone, so the result feels mechanical. Voila skips the text middleman and learns a direct path from audio in to audio out. That's what lets it hit 195 ms, faster than the roughly 200 to 300 ms it takes a person to answer.
The trade-off is where you use it. Voila ships as open weights and a hosted demo, so there's no polished consumer app, no monthly plan and no support desk. If you want it in a product, you or your team handle the setup.
Getting Started
- Open the Voila demo on Hugging Face to try live voice chat, text-to-speech and speech recognition without installing anything
- Download the open weights from Hugging Face if you plan to run it on your own hardware
- Follow the GitHub repo to set up the Python environment and load a Voila checkpoint
- Write a short text instruction to set the speaker's persona and tone, or add a 10-second audio clip to clone a voice
- Run inference for a task like spoken dialogue, transcription or speech synthesis, then fine-tune if you need translation
Product Information
A quick look at Voila's pricing, supported platforms, and performance.
Best for
The users, tasks, and scenarios where this tool fits best.
Users
- AI researchers and grad students
- ML engineers building voice products
- Indie developers and hobbyists
- Voice and audio designers
Tasks
- Real-time spoken dialogue
- Voice cloning
- Text-to-speech
- Speech recognition
- Role-play and character voices
Scenarios
- Research on speech models
- Prototyping a voice agent
- Local and offline use
- Multilingual demos
- Teaching and workshops
Key features
Full-Duplex Conversation
Voila can listen and speak at the same time, which is what separates it from turn-based assistants. The autonomous model processes your audio stream and its own output in parallel, so it can backchannel, take an interruption or overlap speech the way people do. In practice, that turns a rigid question-and-answer loop into something closer to a real call. The company's own framing is that it wants the model to react and to initiate, not just wait.
195 Millisecond Response Latency
The headline number is 195 ms from end of speech to start of reply. For context, people typically answer in about 200 to 300 ms, so Voila lands at or under human speed. That matters because delay is what makes a voice assistant feel robotic. The low figure comes from dropping the text pipeline and modeling audio directly, which cuts the pauses that stack up when three separate modules hand work to each other.
Voice Customization and Persona Control
You shape a voice two ways. A written instruction sets the persona, tone and manner, much like a system prompt. A separate audio clip teaches the model a speaker's timbre, and the clip can run from a few seconds to hours. The team used that shortcut to build more than a million ready-made voices, so you can also pick an existing one instead of training your own. A 10-second sample is enough to start.
One Model for Many Audio Tasks
The same model handles dialogue, text-to-speech and speech recognition without a separate build for each. Speech translation is possible with light fine-tuning rather than a full rebuild. For developers, that means one checkpoint and one codebase to maintain instead of gluing together three services. It is the kind of consolidation that saves on infrastructure and on the time spent keeping pieces in sync.
Open Weights and Code
Everything is open. The weights and code sit on Hugging Face and GitHub under the Maitrix.org organization, and anyone can download, run and adapt them. That is the main draw for research, since you can inspect how the system works instead of treating it as a black box. It also means there's no per-minute bill and no vendor that can change the terms under you.
Multilingual Support
Voila is trained on multilingual text and audio and covers six languages: English, Chinese, French, German, Japanese and Korean. It handles both understanding and generation in those languages, so a conversation can move between them. For teams serving more than one market, that breadth avoids standing up a separate model per region.
Pros and cons
Pros
- 195 ms latency puts replies at or below human response speed, which is what makes a voice assistant feel live
- Full-duplex design lets the model listen and speak at once, so interruptions and backchanneling work naturally
- Open weights and code mean no license fee and full freedom to inspect, retrain and ship
- Over a million built-in voices plus 10-second cloning make custom characters quick to set up
- One checkpoint covers dialogue, text-to-speech, speech recognition and, with tuning, translation
Cons
- There's no consumer app or hosted service, so you handle setup, hosting and maintenance yourself
- Running the model locally needs real compute, which puts it out of reach for casual users on a laptop
- The six-language list is narrower than commercial platforms that advertise dozens of languages
- Translation needs fine-tuning rather than working out of the box, so it adds work before it pays off
Frequently asked questions
Voila is an open-source family of voice-language models built by Maitrix.org with UC San Diego and MBZUAI. It runs real-time voice conversations end to end, from audio in to audio out, and supports spoken dialogue, text-to-speech, speech recognition and, with tuning, speech translation.
Related content
Explore related tools, skills, and articles for Voila.
Voila Alternatives
Foley
Hanamori Labs, LLC · Voice & Language · Productivity · AI AssistantFoley is a voice to-do list that turns a spoken sentence into a filed task, complete with project, date and time, then helps you plan the day around what is already there.
HireKey
HireKey · Voice & Language · Productivity · LeaningHireKey is an AI job search copilot built by a former tech recruiter. It checks role fit, tailors resumes, runs interview practice and tracks every application.
Fluence
Yay Software (Yannick Huchard) · Writing · Voice & Language · ProductivityFluence is an Android app that reads PDFs, EPUBs, Word files and web articles aloud, follows each sentence as it speaks, and answers questions about the text on the phone itself.