UtteraTranscribe audio
Upload a recording, get the text. It can tell the speakers apart, add timestamps and, if you ask, produce the summary and the analysis in the same pass.
One engine, two doors. A set of web apps where the work gets done without writing any code —the Studio and, as separate apps, Pronounce, Sounds and Music—, waiting for you the moment you sign up. And an API so your own software does it on its own, with nobody opening a browser. Below is each one, one paragraph each, with a link to its manual.
One page with four tabs. Drop a file in or paste some text, press the button, download the result. It is what you will see when you walk in.
Upload a recording, get the text. It can tell the speakers apart, add timestamps and, if you ask, produce the summary and the analysis in the same pass.
Paste a text and hear it. Nine languages in the standard voice and around thirty in the high-quality one, at the speed and in the format you choose.
Upload a short sample of a voice and the text is read in that voice. Meant for dubbing yourself, not for impersonating someone else: that is forbidden, and written down.
A recording goes in one language and comes out in another. Tick the box that keeps the original voice and it comes out spoken by that same voice instead of by a stock one.
From a long meeting or call you get, up front, a summary —what was agreed and what's left to do— to grasp in seconds what it was about, with the full transcript below for the detail.
Turn a meeting or call into a report where each statement is anchored to its quote, speaker and minute, as a signed PDF a third party can check.
The Studio shows the estimate before you launch the job and the real figure after, in credits. No surprises at the end of the month, because there is no end of the month: it is on the same screen.
There are real limits —a three-word sentence translates worse than a fifteen-word one, and silence can make the engine invent— and they are written down instead of being discovered by surprise.
The second application, and deliberately narrow: three steps, with no options in the way. Same menu, same account.
Write the sentence you need to say tomorrow in the language you already speak, and it comes back in the one you are learning, with its phonetic transcription underneath.
The audio is generated when you ask for it, not before, at whatever speed you need. And you can play it again as often as you like without paying for it twice.
Record your attempt and it is compared sound by sound against the right one: what you did, what you should have done, and in which word. If you also want it explained in plain words, a button asks for that — and warns you it costs more before it runs.
The third application, and the first that does not work with voice: you describe a sound in words and out comes the clip. Same menu, same account.
«A wooden door slamming shut». That is all: you choose how long it runs and it comes out as 48 kHz stereo WAV, the rate video works at.
You describe what happens —«a motorcycle stops after eight seconds, the rider gets off and walks to the sea»— and it is split into events placed in time. You see the plan, you fix it, and only then are they generated and mixed.
Two steps and not one because six events are six generations: spending them on a plan that wasn't what you meant is throwing them away.
Every clip comes back with its seed. The same description with the same seed gives the same sound, so the one that came out right is not lost.
Another separate app: you describe the music you need and out comes a finished piece, up to six minutes and twenty seconds. These are not loops to assemble.
«Tribal African techno with djembe and deep toms, 128 BPM». Naming the style, the instruments and the tempo helps far more than asking for «something upbeat».
How long it runs and how much quality. Both change what it costs, so the application shows you the credits before generating: moving a control is never a surprise.
And the piece is built whole for the length you ask — a three-minute piece is not a thirty-second one repeated six times.
What you generate is yours, with commercial use from the Startup plan up. No catalogue to dig through and no annual licence behind it.
Three apps built for the phone. They open in the browser, use the microphone and can be installed like an app —no store, no download.
Live two-way voice translation. You speak and it plays back, translated, in the other language. One button per speaker: ideal for travel and face-to-face conversations.
Your pronunciation coach. Pick a language, type a phrase, listen to it and record yourself; we give you a score and sound-by-sound feedback, with phonetics, to fix your accent.
Transcribe, summarise and report on your recordings. Pick a file, record on the spot or share a voice note from another app; you get the text, a summary or a PDF report (with an optional evidence graph).
They are web apps (PWAs): from your phone's browser, "Add to Home Screen" and they sit there as one more icon, full-screen. Nothing to install from a store; they update themselves.
The services, over an audio file or a text. We speak the same dialect as OpenAI's API, so in many cases changing the base URL is enough.
Audio in, text out. With speaker separation, timestamps, and the option to ask for the summary and the analysis in the same request, without uploading the file again.
Three kinds of voice, not two: standard, high quality, and cloned from a sample of your own. And you can start hearing it while it is still being generated instead of waiting for the whole file.
Text to text, or a whole recording from one language into another. And with transposition: the translation spoken by the voice of the original recording.
How much each person spoke, in what tone, at what pace.
With a limit we impose ourselves: these are acoustic estimates, they get things wrong, and they must not be used to make decisions about a person. They are there to improve a process, not to judge anyone.
An hour of audio comes back as what was agreed and what is still pending, structured so you can map it onto fields.
And you pay your LLM less along the way: 15,410 tokens of transcript against fewer than 700 of summary, measured on a real recording.
Send the sentence someone meant to say plus the audio of how they said it, and it returns which sound was made, which one was due, and in which word. It is what drives the application above, and it is open for your own language platform to use.
Sign up, create a key, send it in a header. A key can be restricted to specific IP addresses, so a leaked one is of no use from anywhere else.
There is published, tested code for Asterisk phone systems, for n8n and for AI agents. They are not toy examples: they are in production.
A free account with credits to try it out, no card. You are charged for what you consume —per second of audio, per character of text—, not per request, and every response carries a header with exactly what was billed.
Create a free account Read the documentation
Processed entirely in Spain · We do not keep your audio · We do not train on your data · Open source