What we are learning as we build this. No press releases: bugs that cost us a day, measurements that came out the opposite of what we expected, and decisions that need explaining.
Most AI summaries ask you to trust them. Ours doesn't: every claim is anchored to its quote, its speaker and its minute, with its status, and the PDF is signed so a third party can check it without trusting us. What the evidence graph is, and how to request it over the API.
A consultation, a statement, a client meeting. Two sectors that live on recorded conversations —healthcare and law firms— and share two demands: audio that is genuinely confidential, and being able to trace who said what. What Uttera offers for that, and what it does NOT.
Whisper is OpenAI's speech recognition model, published under an open licence, and it's what sits underneath almost everything that transcribes today. Here's what it is, how it works, when it's worth self-hosting, and what its limits are.
Voice cloning is building a synthetic voice that reproduces the timbre of another from a short sample. Here is what it actually is, how it learns from that sample, when it sounds native and when it carries an accent, and whose voice you may clone.
Text-to-speech turns written text into spoken audio. Here's what it is, how it works under the hood without the jargon, and what to actually check —quality, languages, latency and cost— before you pick a provider.
Speech-to-text turns spoken audio into written text. Here's what it is, how it works under the hood without the jargon, and what to actually check —accuracy, languages, latency and cost— before you pick a provider.
Diarization answers one question: who speaks and when. It doesn't tell you who anyone is; it separates voices and labels them. Here's what it is, how it works, when you need it, and where its limits are.
MCP is the standard way to give an AI assistant tools. Instead of describing every API by hand, you give it one address and it discovers on its own what it can do. Here is what it is, how it works, and when to build one.
A watermark is an imperceptible signal embedded inside the sound that says "a machine made this". Here is what it is, why the provenance of audio matters, and why a signature is mainly for ruling out, not for boasting.
Which services each voice API offers —TTS, transcription, diarization, cloning, translation— and what it charges. An honest, dated comparison so you decide on data, not marketing.
Truly sustainable AI: Uttera doesn't run on anyone's cloud: the engines live in our own facilities on self-generated solar power. Electricity bill at zero since the service opened on 15 September 2026, measured at the utility company's own meter and verifiable live.
What really separates providers isn't where they say they process, it's what they do when they fill up. We answer 503. What we don't do is forward your audio to somebody else's API so you don't notice.
Pulling an effect out of someone else's video is a licence you don't have, at the wrong length, with someone else's room in it. Generating one takes seconds and comes out to measure — with its limits, which we spell out too.
Two stacked licences, a revenue threshold, a mandatory attribution and a training-data fact almost nobody mentions. What you need to know before putting a generated track into something you publish.
Article 50 has applied since August 2026 and our deadline is December. We have complied since September, using a public tool from Meta — so you can check it yourself, without asking us for anything.
You write the sentence you actually need, hear it, say it yourself, and we give you back what we understood, word by word. No score, no marks and no gamification: an honest mirror.
Every report carries two signatures and each answers a different question: whether the text came from us, and whether the file is exactly the one we issued. You check by uploading the PDF, with no account and without trusting anyone.
Give it empty audio and it doesn't return an empty string: it invents a sentence. And the worst part isn't paying for it, it's your automation acting on it.
It started as a tool for internal use in a Spanish family business group. It wasn't designed to be sold. That's why it's built the way it is.
Billing is per second of audio, not per request. That changes how it pays to split the work, and there's a trick that saves you money before you send anything.
The difference between "separating speakers" and "identifying people" takes two paragraphs to explain and saves an expensive disappointment.
The short sentence said more than we meant, and it closed a legitimate door on us. We went through all five places it appeared on the site.
Your phone system is already writing .wav files somewhere. Pointing a script at that folder is half an hour. Putting a voice inside the call is a day, and here are the five traps.
They are acoustic estimates. They get things wrong. And there's one use we won't support even if someone pays for it.
We keep the audio we generate for one hour, and it costs 10%. An hour of cache is not an hour of access to your data, and you can switch it off on every request.
A 70-minute recording is 15,410 tokens of transcript and under 700 of summary. Twenty to thirty times less, and the recording never leaves here.
The engines that move Uttera are open source. This is what it takes to stand them up on your own machine, without going through us.
It isn't a quality decision. It's about how each one uses GPU memory, and choosing wrong is the fastest way to lose an afternoon.
Ten scripts and an instructions file. The hard part isn't the API: it's telling the agent when to use each thing and what not to believe.
Same routes, same parameters, same voice names. What you change is the base URL, the key… and one more thing that isn't obvious.
It isn't a marketing label. It's the difference between a procurement file that passes and one that sits waiting for a legal opinion.
The technical part is solved with ten seconds of clean audio. The part almost nobody writes down is the other one: that a voice is an attribute of a person.
Uttera