Linden by Speechmatics

Linden by Speechmatics is a transcription tool. Linden is Speechmatics' speech-to-text model for AI voice agents, delivering final transcripts in about 350ms with speaker diarization, 55+ languages and pricing from $0.30/hour, usable via LiveKit, Pipecat or a direct API.

Linden is the speech-to-text model behind Speechmatics' Agent STT, an API aimed at teams building AI voice agents rather than general transcription. The pitch is narrow and practical: return final transcripts quickly, label which speaker said what, and avoid the errors that derail an agent, such as misheard names, numbers and other details a caller expects to be handled correctly. The service reports final transcripts about 350ms after speech ends, which leaves the language model more time to draft a reply before a caller notices a pause. Speaker diarization and turn signals (start and end of speech, start and end of turn) are included by default, so an agent can respond to the right voice in a multi-speaker call. Turn detection currently relies on silence, with server-side semantic detection listed as upcoming. Custom vocabulary is supported, and Linden is available across more than 55 languages, though full multilingual support is described as still coming. In practice there are three ways in. Linden can be selected inside LiveKit Inference with a config change or a toggle, needing no new account or SDK. Developers who want more control can use the pre-built integrations for LiveKit Agents or Pipecat. The most hands-on route is connecting directly to the /v2/agent endpoint with an existing Speechmatics API key and the linden-1 model, then assembling a custom STT, LLM and TTS pipeline. A Python client streams microphone audio and fires callbacks on transcript segments and end-of-turn events. Pricing starts at $0.30 per hour, with volume discounts reaching $0.15 per hour, and new accounts begin with $100 in free credit. Compared with general-purpose transcription APIs, Agent STT trades breadth for a focus on conversational latency and speaker attribution. It suits developers of phone agents, contact-centre automation and other real-time voice products, while teams needing batch transcription or fully mature multilingual turn detection should check the current feature status first.