Today, we're excited to announce the launch of Rime Speech-to-Text. So excited in fact, that we're giving it away for free for 30 days!
Customers have asked us for this since the first Rime voice shipped. We held off until we could build it from the thing that makes speech-to-text great on a live phone call, which is years of unscripted conversation, written down by hand.
Rime’s secret sauce? How we collect voices. We built a recording studio, invited people in to talk to each other without a script, and then did the part nobody puts in a launch post: our linguists wrote it all down by hand, under one set of conventions, so a false start, a repeated word, or a number said digit by digit is written the same way every time. This is the magic of why Rime voices feel human, this unique dataset of full duplex conversations.
That is the material Rime STT was built from.
What you control with Rime STT
We make models for builders who want control over what their agent says. Speech-to-text gets the same treatment. The decisions most vendors make for you are settings here, and every one of them is set per request, so one agent can transcribe a payment confirmation one way and an intake interview another.
How the text arrives
- Rewrite as it goes (the default): every update is the whole transcript so far, so earlier words can be corrected as the sentence completes. "Flight to four seven" becomes "Flight 247" the moment the number ends, and your agent already has it. Use it when the agent acts on the live text.
- Add as it goes: each five-second window is transcribed once and appended, and nothing changes after it appears. Use it when the text is going somewhere that can't take a correction, like a screen, a form, or a log. Most STT vendors give you just one of these, but we decided to put you in control of how text arrives.
What the text keeps
- Clean (the default): readable text, digits for numbers, fillers removed. "Flight 247 is now boarding." This is what you want in a CRM note, a log, or a confirmation read back to the caller.
- Word for word: everything, in the order it was said, numbers as words. "uh flight two four seven is uh now boarding." Verbatim transcription isn't new. What it's for is. "I want to, um, cancel? Or can I pause it?" is a person who hasn't decided, and your model can only notice that if the hesitation reaches it in the transcript.
What it listens for
- Custom terms: the provider names, medications, products, or menu items your callers will say, sent with each request, up to 1,024 characters. They're hints: a term won't appear where nobody said it, and a name that isn't on the list can still come out right.
- Got one wrong? Click the word in the dashboard transcript, type the right spelling, and it goes on the list.
How fast is Rime’s STT?
Try it on the words your callers actually say
Open the dashboard, press the mic, and say an order number, a provider's name, or your company's name the way a caller would. Then say it badly. Switch between clean and word for word and run the same clip again; the words that changed are highlighted. When the transcript looks right, "Use this in your code" gives you a snippet with the same settings.
Rime STT is free for 30 days (until November 19th) - sign up for a free account to start using it now. Already a customer? Check out our docs to get started.
More from the lab
Fine-tuning voice AI: why your call data is the moat
A caller orders "a bowl, no pico, sub queso, extra rice on the side." A generic voice agent catches "bowl" and "rice" and guesses at the rest. The one trained on ten thousand of that restaurant's calls knows exactly what "sub queso" means at that counter, because it has heard people say it every day.
