Transcript (en)

So for speech-to-text, a voice agent actually has to do three things. Firstly, listen to the user and transcribe what they say. Secondly, work out when they're actually done. And finally, and most crucially, hand off to an LLM so they can respond. Now, all of this might sound really easy on paper, but it can actually be quite hard to get it exactly the way that you want. So today, I want to show you something we've been working on, a robust API built from the ground up to make working with voice agents easier than ever. Introducing Speechmatics Agent STT, powered by our new real-time model, Linden1. Agent STT is our take on what a voice API should look like for voice agent builders. Now we handle all the fine tuning for you and give it back to you in three new message types designed specifically for your voice agent stack. Here's a demo that transcribes using our Agent STT API. Right now, only our transcription content messages are being shown. You'll see partials turning up in gray and finals in a white text replacing those when I stop talking. This might seem very similar to our real-time API, but instead of returning individual words, we return full segments of speech. That means no more splicing words and punctuation together, we give the entire segment start to finish to you Next up we have turn signals which indicate when a user has started or finished talking This is the stuff you can wire up to an LLM to tell them you interrupting or handing off So if I talk now you see these lines indicating boundaries for user speech. Turn signals are aware of language, hesitations and whether or not the sentence actually sounds complete, not just whether there's silence. These are all the signals you can use to interrupt or hand control to an LLM. And the great part about it, it's all handled under the hood. So you don't have to manually synchronize up transcript timestamps or turn detection into your pipeline. Finally, we have our speech signals, which are purely acoustic indicators. This is our server-side VAD talking, which fires off the instant sound starts or stops looking like speech. This is great for UI indications of when the user is speaking, for example, a waveform, form, a badge, or in this case, a little display at the top of the page that lights up and provides reactive visuals when the user is speaking. These are purely acoustic and don't go off of the same sentence context that turn signals do. AgentSTT is in preview today. To access it, you'll need a Speechmatics API key and point it at the preview endpoint we showed you earlier. As always, the full docs are linked in the description below, and we'd love to hear from you if you're building on it.