Speechmatics & LiveKit Inference target the accuracy gap that breaks voice agents in production

As of today, developers building voice agents on LiveKit Inference can switch to Speechmatics in under a minute, through the UI or in code. Linden, our first model built specifically for voice agents, is now available through LiveKit's platform for using leading voice AI models from one place. That means no separate API key, account, or invoice.

In a demo, a voice agent typically hears one speaker, clean audio, and a short exchange, the conditions where every speech-to-text provider looks good. Production is rarely that tidy: a caller with a strong accent, a noisy cafe, a 16-digit card number read out over a bad line. When the transcript breaks on the words that matter, everything downstream breaks with it.

Together, we solve that for the builder. Linden gets the words right where it's hardest: across 55+ languages, non-native speakers and strong accents, and alphanumerics such as account and card numbers read out over a bad line. LiveKit Inference handles the pipeline, the routing, the billing, and the turn detection that decides when the agent speaks. Adopting a speech provider used to take weeks. Now it's something you can simply select.

What separates an agent people trust from one they find exhausting is whether it can understand what you said. Our new model solves for this. Linden not only excels at 55+ languages, but also transcribing non-native speakers, dialects, and accents. LiveKit built the infrastructure that made those deployments possible, and its community has worked out what production voice AI actually takes, in the real world, faster than anyone else.
— Ricardo Herreros-Symons, CRO, Speechmatics

Proven in production

Norwegian conversational AI company boost.ai already runs Speechmatics and LiveKit in production across the voice agents it builds for European enterprises, on infrastructure that also powers xAI Grok and Salesforce Agentforce.

Voice agents live or die on the STT layer. Speechmatics gives us what actually matters in production: low-latency real-time models, fast iteration on the hard edges like alphanumerics and end-of-utterance, and serious language breadth across Europe. Just as importantly, the people behind it are a pleasure to work with — close, hands-on, and genuinely invested in getting things right.
— Rasmus Hauch, CTO, boost.ai

How it works

LiveKit Inference, launched in October 2025, lets developers build a voice agent in minutes. A single API key calls the most popular STT, LLM, and TTS models, and models swap with a one-line code change. Consolidated billing, rate-limit management, and usage reporting all run through LiveKit Cloud.

Speechmatics is available natively through LiveKit Inference, via the Speechmatics plugin for the LiveKit Agents open source framework, or through LiveKit's Agent Builder.

LiveKit Inference exists so developers can ship and iterate fast, without the headache of managing API keys or rewriting their stack every time a model changes. Adding Speechmatics equips voice agents with exceptional speaker diarization that clearly tracks who said what with any number of speakers from day one. This capability is especially powerful for physical AI and real-world multi-speaker environments.
— David Zhao, CTO & Co-founder, LiveKit