Your own voice AI: speech, transcripts and phone agents on your servers.
You pay a voice service for every character it speaks or every minute it listens. We set up open-source speech models on your own hardware, connect them to your product or your phone system, and test them in your language before you switch.
Speaking, listening and the call, on hardware you own.
Voices you choose by ear
The same sentences from several models, in your language.
Licences read twice
Once for the code and once for the trained voice files.
Calls at the same time
Measured on your machine, so you know its limit before you rely on it.
Reply delay timed
Each step, from the caller's last word to the agent's first.
Your phone system
Connected to the provider and the equipment you already use, where the framework supports it.
The product around it
Agent rules, hand-over to a person, call records and reports.
Voice products pay in three places: for speech the product generates, for audio it turns into text, and for the service that runs the whole phone call. All three are billed by use, so a product that answers more calls pays more every month.
Open projects cover each part, and our directory lists them: alternatives to ElevenLabs for generated speech, alternatives to Deepgram for speech to text, and alternatives to Vapi for phone agents.
The three parts
Speaking. A text-to-speech model turns your product's text into audio. The directory lists Chatterbox, Piper and Kokoro FastAPI among others. The code and the trained voice files can carry separate licences, and the directory marks the models whose voice files restrict commercial use, so this is checked first.
Listening. A speech-to-text model turns the caller's audio into text. The directory lists Whisper, WhisperX and Parakeet among others.
The conversation. A framework joins listening, a language model and speaking into a live call. Pipecat, by its own description, wires the three together over a phone line, and the directory lists LiveKit Agents for the same job.
One project, Speaches, describes itself as presenting OpenAI's audio request format on your own hardware for both speaking and listening, which lets tools written for that format call it after an address change.
What we do
Test voices in your language. Quality differs by language and by voice. We generate the same sentences from your product with a short list of models, and you listen and choose.
Size the hardware for calls at the same time. The question is how many calls your machine carries at once before replies slow down. We measure that on your hardware and give you the figure.
Connect it to your phone system or your product. Including the phone line provider you already use, where the framework supports it.
Measure the delay. In a phone call, the time between the caller finishing and the agent starting to speak decides whether the call feels natural. We time each step and show you where the time goes.
Build the product around it. The rules for what the agent may do, the hand-over to a person, call records, reports. This is the larger part of the work.
Voice cloning
We clone a voice only with the written consent of the person whose voice it is, and only with a model whose licence allows your use.
Common questions
Can we run text to speech on our own server in place of ElevenLabs?
Yes. Our directory lists open text-to-speech projects such as Chatterbox, Piper and Kokoro FastAPI on its alternatives to ElevenLabs page, with the licence of each. Reveneau tests a short list in your language on your own sentences, sets up the one you choose on your hardware, and connects it to your product.
Which open voice model sounds best in my language?
There is no general answer, because quality differs by language and by voice. Each voice model's own page lists the languages it supports. Reveneau generates the same sentences from your product with several models in your language, and you choose by listening. We make no claim that an open voice model matches a paid voice until you have heard both.
How many phone calls can one GPU handle at the same time?
It depends on the models, the GPU and how fast replies must be, and many published figures come from companies that sell hardware or voice services. Reveneau measures it on your own machine: we add calls at the same time until the reply delay passes the limit you set, and that count is the capacity of your machine.
Can an AI phone agent run fully on our own hardware?
Yes. A phone agent has three parts: speech to text, a language model, and text to speech. Open projects exist for each, and a framework such as Pipecat or LiveKit Agents joins them into a live call. All three parts can run on your machines, or one part can stay with a paid service while the others move.
Can our own voice agent connect to our existing phone system?
Yes, where the framework supports your phone line provider. Pipecat's own description says it works over a phone line and includes support for six carriers. Reveneau checks your provider against the framework's list at the start, and writes the connection when your provider is missing from that list.
Can you clone a voice for our product?
Reveneau clones a voice only with the written consent of the person whose voice it is, and only with a voice model whose licence allows your use. Some open voice models restrict commercial use of their trained files, and our directory marks which. The rules on voice consent differ by country, so check them with your lawyer for the places you operate.
Why does reply delay matter in a voice agent?
A phone call feels slow when there is a silence after the caller stops speaking. That reply delay is the sum of three steps: turning speech into text, the model deciding what to say, and generating the audio. Reveneau times each step on your hardware and works on the slowest one first.
Can we move only one part of our voice product and keep paying for the rest?
Yes. The three parts are separate, so you can move generated speech to your own server and keep a paid service for speech to text, or the other way round. Reveneau builds the product so each part can be changed without rebuilding the others, which also lets you return to a paid service if you need to.
Other services
Replace a paid AI service with one you own.
You pay an AI vendor by the month, by the seat or by use, and the bill grows with your business. We set up an open-source replacement on your own servers, connect it to your product, and build the software around it. At the end the servers, the data and the code are yours.
Your own AI model server, in place of a bill for every request.
You pay a model vendor for every request your product makes. We set up an open language model on your own machines, give it the request format your code already uses, and test it on your real work before you switch.
Your own company chat assistant, in place of a seat for every person.
You pay a monthly seat for every person who uses a chat assistant. We set up an open-source chat assistant on your own servers, with your company sign-in, and connect it to the models you choose.
Your own search over company documents, on servers you control.
You pay for a tool that searches your company's files and answers questions from them. We set up an open-source replacement on your own servers, connect it to where your documents live, and keep each person's view limited to what they are allowed to see.
Your own automation platform, in place of a bill for every task.
You pay an automation service for every task it runs, and the automations your business depends on live in someone else's account. We set up an open-source automation platform on your own servers and rebuild your automations on it.
Your own AI coding tools, with your code kept on your machines.
You pay a seat for every engineer who uses an AI coding assistant, and parts of your source code are sent to the vendor with each request. We set up open-source coding tools for your team, connected to a model you choose, including one on your own servers.
Build with us
Release software with confidence.
Move faster without lowering your standards. A software development partner whose senior teams integrate into how you decide and stay accountable through delivery.
An MVP development company built for your next milestone.
An MVP development company for founders: senior judgment and focused sprints that strengthen the business, not just release features.
A build partner for the companies you believe in.
Technical due diligence and senior leaders who reduce the technical risks across your portfolio.