Speechmatics targets voice agent accuracy gap with Agent STT

Cambridge-based Speechmatics launches a speech-to-text API built to catch the small errors that silently break enterprise voice agents.

By Disrupts Editorial Team Sun, 27 September 2026 AI level 2 of 5

AI level 2 of 5: drafted by our AI editorial assistant from the company's announcement; fact-checked, edited and signed off by Mark Walker, Editorial Director. What the levels mean

White robotic arms with blue details operate on metallic components moving along a conveyor belt in a brightly lit industrial assembly line.

Speechmatics, the Cambridge voice AI company, has launched Agent STT, a speech-to-text API engineered specifically for the failure modes that undermine production voice agents. Powered by a new model called Linden, the product targets a category of error that aggregate accuracy scores routinely conceal: a misheard digit in an account number, a dropped negation, or a missed single-word confirmation. Each of these can propagate silently through an otherwise correctly functioning agent stack and produce a wrong outcome at the point of action.

The launch sits at the intersection of two fast-moving infrastructure layers: the proliferation of large language model-driven voice agents across enterprise workflows, and the growing recognition that the speech-to-text layer underneath those agents has become the reliability bottleneck rather than the reasoning layer above it.

Precision over averages

Speechmatics' framing positions Agent STT around semantic accuracy rather than raw word-error rate, the metric that has historically dominated speech recognition benchmarking. In the Pipecat open-source STT benchmark, which tests streaming models across a range of conversational scenarios, Linden recorded a 1.05% pooled semantic error rate with a 369 millisecond median finalisation time. Of 23 streaming models evaluated, none was both faster and more accurate, placing Linden on what the benchmark describes as the speed-accuracy Pareto frontier.

Mark Backman, VP Product at Daily, which maintains Pipecat and the benchmark, noted: "Semantic error rate is central to the open source benchmark we maintain. Speechmatics Agent STT powered by Linden gives developers and enterprises another strong option for the speech layer."

The product supports more than 55 languages, finalises spoken segments in under 350 milliseconds in Speechmatics' internal testing, and includes custom vocabulary of up to 1,000 terms, live speaker diarisation, Speaker ID, and conversational event signals alongside transcripts. Developers can integrate via direct API or through Pipecat and LiveKit, meaning the speech layer can be dropped into existing agent stacks without rebuilding surrounding infrastructure. Launch pricing starts at $0.30 per hour, falling to $0.16 per hour at volume.

The infrastructure play behind the product

The strategic read-across here extends well beyond a single API launch. Enterprise adoption of voice agents has accelerated sharply across sectors where a transcription error carries regulatory or financial consequence: banking, insurance, healthcare administration, and customer operations in telecommunications. In each of these verticals, a misheard account number or a failed negation is not a user-experience inconvenience but a compliance or settlement risk. That consequence structure is precisely what elevates speech-to-text from a commodity input to a critical-path infrastructure decision.

Ricardo Herreros-Symons, Chief Strategy and Revenue Officer at Speechmatics, identified accuracy as the deciding variable once a baseline latency threshold is met: "However advanced the agent, if it misses a digit, a name or a single 'not', the whole interaction can go wrong." That framing reflects a broader pattern in the voice AI stack: as reasoning models improve rapidly and latency normalises, the competition for enterprise workloads is shifting to the reliability and auditability of the underlying perception layer.

From a capital-allocation perspective, the voice AI infrastructure segment is attracting increasing attention from both corporate venture arms and specialist deep-tech funds, drawn by the combination of high switching costs once a speech layer is embedded in production and the growing volume of regulated-industry workflows being automated. Speechmatics, privately held and Cambridge-based, competes in this space with US-centric players including Deepgram and AssemblyAI, as well as the speech APIs offered by hyperscalers. The decision to benchmark publicly against an open-source standard rather than proprietary tests signals a deliberate move to establish credibility with enterprise procurement teams who are increasingly sceptical of vendor-reported accuracy figures. For cross-sector investors watching the agentic AI buildout, the lesson is the same one emerging across the stack: the value in the automation wave may accrue less to the visible reasoning layer and more to the unglamorous infrastructure underneath it.