Speech-to-text for AI voice agents

Agent STT runs on Linden, a model purpose-built for AI voice agents. It accurately captures what was said and by who in 350ms. From $0.30/hour.

Start with $100 in free credit.

Use it through LiveKit, Pipecat, or the API directly.

Case study

AI Media | Case study

Delivering 120X more with voice AI

Powering live content through AI-powered transcription, built on industry-leading voice AI

Case study

LiveKit | Case study

Enabling 100,000+ developers with leading speech recognition

Pairing LiveKit’s flexible agent framework with Speechmatics to build world-class agents

Case study

Adobe | Case study

Cloud-grade speech recognition on-device for Adobe Premiere

Run the most accurate on-device transcription locally; efficient enough for a laptop, powerful enough for professional work.

Case study

NCI | Case study

Redefining real-time captioning

How NCI delivered a 99% increase in usage of automated captioning

Case study

Media Track | Case study

Delivering a 20% leap in accuracy improvements

Improved transcription performance across more than 20 languages for their global clients

Case study

Prosodica | Case study

Driving better conversations at scale

Leveraging speech recognition to track customer interactions, highlight key insights, and raise contact center performance

Case study

Stenograph | Case study

Industry-first on-device speech recognition for CATalyst VP

How Stenograph brought industry-first, on-device speech recognition into CATalyst VP for real-time legal voice reporting.

Built for what AI voice agents need

Not a general-purpose transcript. Accuracy, speed, and signal, tuned for the moments a voice AI agent can’t afford to get wrong.

Accuracy

Catches the details that matter

Linden transcribes 55+ languages and catches names, numbers and details that general models miss, so your agent acts on what was actually said.

Languages

Latency

Final transcripts in 350ms

End-of-speech finalization lands in 350ms, giving your LLM more time to draft a reply before callers notice the pause in the conversation.

Latency docs

Speaker focus

Knows exactly who is talking

Speaker diarization attributes every word to the person who said it, so your agent acts on the right voice, not the loudest one, in multi-speaker calls.

Speaker diarization docs

Price

$0.30 an hour to start

Linden costs $0.30/hour, with volume discounts that bring the price down to $0.15/hour as your usage scales, and no change to accuracy or latency.

Full pricing

See turn detection in action

Toggle partial transcripts and turn signals to see exactly when Agent STT knows a caller has finished speaking.

Understand the signalsShow examplesHide examples

Partials Turn signals

Your voice starts the conversation.Start the microphone to see words and turns appear.

How to build with Agent STT

Agent STT drops into the voice agent stack you already use, or straight into your own pipeline. Pick the level of control you need.

Path	Best for	How it works
LiveKit Inference	Fastest to launch	Select Linden in LiveKit Inference. One config change or a toggle in the LiveKit UI, no new account or SDK.
LiveKit Agents & Pipecat	Full control in code	Open-source framework, pre-built Agent STT integration for LiveKit Agents and Pipecat. Proven pattern, no build from scratch.
Direct API	Maximum control	Connect to /v2/agent with your existing Speechmatics API key and model linden-1. Build your own STT, LLM, and TTS pipeline.

Build AI voice agents faster with a flexible API

Build AI voice agents without rebuilding the speech layer from scratch. Speechmatics gives teams a flexible API for accurate, low-latency transcription. Developers can focus on the agent experience, workflow logic, and what happens after the conversation.

Or build on the voice agent frameworks you already use, like Pipecat and LiveKit.

Python

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

# pip install speechmatics-agent-stt pyaudio

import asyncio

from speechmatics.agent_stt import AgentSttAsyncClient, Microphone, ServerMessageType


async def main():
    client = AgentSttAsyncClient()  # reads SPEECHMATICS_API_KEY
    mic = Microphone(sample_rate=16000, chunk_size=1024)

    @client.on(ServerMessageType.ADD_SEGMENT)
    def on_segment(message):
        print(message["segment"]["transcript"])

    @client.on(ServerMessageType.END_OF_TURN)
    def on_turn_end(message):
        print("-- speaker finished, your agent can reply --")

    mic.start()
    async with client:
        while True:
            await client.send_audio(await mic.read(1024))


asyncio.run(main())

Speaker 1

Hi, yeah, I think I've lost my card. Or it might've been stolen, I'm not totally sure.

Agent

Sorry to hear that. I can help freeze the card right away. Can I get your full name and date of birth?

Speaker 1

Sure, it's Daniel Okafor. Date of birth, 14.03.1991

Agent

Thanks, Daniel. I've found your account. Just to confirm, are you reporting the card ending 4482 as lost or stolen?

Linden 1$0.30/hr Speech-to-text for voice agents. Diarization and turn detection out of the box.

What does Linden cost?

Linden is the speech-to-text model behind Agent STT, built for voice agents rather than general transcription. Speaker diarization and turn signals are included by default.

It costs $0.30/hour with volume discounts as usage scales down to $0.15/hour. A calculator for modeling your own usage is coming soon.

Agent STT: Frequently asked questions

Agent STT is a speech-to-text API and model built for voice agents. It runs on Linden, adds custom vocabulary, and tells your agent who's speaking, all in 350ms.

Agent STT runs on Linden today and is available in all 55+ languages. Full multilingual support is coming soon.

Agent STT gives you voice activity and turn signals (start of speech, end of speech, start of turn, end of turn) today, driven by silence detection. Server-side semantic turn detection is coming soon.

Agent STT is a separate package priced and built specifically for voice agents, with turn and speaker signals included. It's not a replacement for general-purpose real-time transcription.

LiveKit and Pipecat today, or connect directly through the API.

$0.30/hour with volume discounts as usage scales.

Data is sent to either our EU, US, or AU endpoints. Customers can also use our Global endpoint which routes you to the nearest location and improves latency.

Yes. Custom Dictionary is fully supported. Add up to 1,000 words to accurately capture place names, acronyms, or any other business-specific vocabulary.

Build voice agents that don't get the details wrong.

Get started with $100 worth of credits.

Speech-to-text for AI voice agents

Delivering 120X more with voice AI

Enabling 100,000+ developers with leading speech recognition

Cloud-grade speech recognition on-device for Adobe Premiere

Redefining real-time captioning

Delivering a 20% leap in accuracy improvements

Driving better conversations at scale

Industry-first on-device speech recognition for CATalyst VP

Built for what AI voice agents need

Built for what AI voice agents need

Catches the details that matter

Final transcripts in 350ms

Knows exactly who is talking

$0.30 an hour to start

See turn detection in action

How to build with Agent STT

How to build with Agent STT

Build AI voice agents faster with a flexible API

What does Linden cost?

Agent STT: Frequently asked questions

What is Agent STT?

Does Agent STT support multiple languages?

Does Agent STT detect end of turn or handle interruptions?

How is Agent STT different from Speechmatics real-time transcription?

Which frameworks does Agent STT work with?

How much does Agent STT cost?

Where is data stored?

Does Agent STT support custom dictionary?

Build voice agents that don't get the details wrong.