Speechmatics

Speechmatics is an enterprise speech-to-text API built by a Cambridge, UK company founded in 2006 by Dr. Tony Robinson, a neural network speech recognition researcher who did his foundational work at Cambridge University. Originally incorporated as Cantab Research Ltd., the company has spent nearly two decades building a single core capability: accurate transcription across every accent, dialect, and language without sacrificing speed. The API powers live broadcast captioning, contact center analytics, medical documentation, and defense intelligence applications where accuracy and data sovereignty cannot be treated as optional extras.

The platform's current model, Ursa 2 (generally available since October 2024), covers 55+ languages in real-time streaming and batch modes, delivers sub-1-second streaming latency with the right configuration, and includes speaker diarization as a core feature rather than a paid add-on. Speechmatics offers three deployment modes: cloud API, on-premises via Docker containers or preconfigured virtual appliances, and on-device edge inference. The company holds ISO/IEC 27001:2022, SOC 2 Type II, GDPR, and HIPAA certifications across its deployment stack. A specialist Medical Model launched in September 2025 achieved 93% general real-world accuracy on clinical audio and 96% medical keyword recall, benchmarked against peer systems from competing vendors. For teams who need the Deepgram or AssemblyAI developer experience without the compliance requirements, see those alternatives; for regulated enterprises and global broadcasters where "good enough" is not enough, Speechmatics is a credible first option.

What Speechmatics produces in April 2026

The Speechmatics Speech API takes audio input (streaming via WebSocket or uploaded batch files) and returns timestamped transcripts, speaker labels, punctuation, and optional confidence scores. The Ursa 2 model, which became the default for all Speechmatics services in October 2024, covers 55 languages in real-time streaming mode, with Irish and Maltese added at launch to complete coverage of all 24 major European languages. In accuracy benchmarks on the FLEURS dataset, Ursa 2 ranked first in 62% of languages tested and top-three in 92%.

Batch transcription rates are available in Standard and Enhanced tiers (Enhanced uses a larger model pass for better accuracy on challenging audio). Real-time streaming uses a max_delay parameter ranging from 0.7 to 4 seconds; the default of 4 seconds maximizes accuracy, while 1.5 seconds is the recommended starting point for voice agent applications. The platform also wraps STT, LLM turn-taking, and TTS into a single product called Flow, a conversational voice agent API available in both cloud and on-premises configurations.

The September 2025 Medical Model deserves a separate note. It set industry benchmarks for English clinical audio with 93% general real-world accuracy (7% WER), 96% medical keyword recall, and a 4% keyword error rate on drug names, diagnoses, and procedure codes, representing 50% fewer medical term errors than the nearest competitor in Speechmatics' own benchmarks. The model is deployed on NVIDIA Triton Inference Server with CUDA acceleration and is available in English, Spanish, French, Dutch, and Finnish, with remaining languages rolling out across the 55-language portfolio.

In March 2026, Speechmatics released what it described as the world's first bilingual Arabic-English production STT model, a single model handling multiple Arabic dialects and English simultaneously. This addressed a long-standing gap flagged in enterprise reviews from the Middle East region.

"We're delighted to work with Speechmatics to drive our live and batch captioning. They continue to be ahead of the pack for all key quality metrics." - Tom Wootton, Product Leader, Red Bee Media, 2024

Where Speechmatics sits versus Deepgram and AssemblyAI

The STT API market in 2026 has three main competitors at the enterprise tier: Speechmatics, Deepgram, and AssemblyAI. The mechanical differences are significant enough that picking the wrong one for your use case has real consequences.

Speechmatics vs. Deepgram Nova-3: Deepgram is a US-based startup with a latency-first architecture. Nova-3 publishes a 5.26% WER benchmark across 2,703 files in 9 domains, and in blended speed-plus-accuracy rankings Deepgram often wins. Its Flux streaming model targets sub-100ms latency for voice agent applications with less configuration overhead than Speechmatics' `max_delay` tuning. Deepgram's Keyterm Prompting allows injecting up to 100 custom vocabulary terms at inference time, while Speechmatics' Custom Dictionary uses a JSON-based approach limited to 6 words per entry with phonetic "sounds_like" alternatives. The critical difference: Deepgram has no on-premises deployment option. For a US SaaS startup building a voice agent, Deepgram's pricing (per-minute rates plus a $4.50/hr bundled Voice Agent API) is more legible at small volumes. For a European broadcaster or hospital that cannot route patient audio through an American cloud provider, Speechmatics is the only option in this tier.

Speechmatics vs. AssemblyAI Universal-3 Pro: AssemblyAI is developer-first and feature-rich. Its LeMUR product layers an LLM directly over audio, enabling summarization, sentiment analysis, Q&A extraction, and entity detection in a single API call without a separate LLM pipeline. AssemblyAI's Universal-3 Pro Streaming claims #1 multilingual accuracy rankings (their own benchmarks) and supports natural language prompting and dynamic key-term injection mid-stream. The mechanical ceiling: AssemblyAI's streaming supports 6 languages (English, Spanish, French, German, Italian, Portuguese). Speechmatics supports 55+ in streaming mode. If your application handles calls from users in 20 countries, AssemblyAI's streaming language cap is a blocker. AssemblyAI also has no on-premises deployment, and add-ons for diarization, medical mode, and LeMUR stack materially onto the base cost. An academic study found that Deepgram trailed both Speechmatics and AssemblyAI on read speech accuracy by statistically significant margins, but the overall picture remains: test on your own audio, because vendor benchmarks do not reproduce consistently across test sets.

"Working with Speechmatics enables us to seamlessly provide our customers with quality, automated speech analytics as part of our solution." - Mariano Tan, President and CEO, Prosodica, 2024

The short version: Deepgram wins on voice-agent latency and US developer ergonomics. AssemblyAI wins on built-in LLM features and developer onboarding. Speechmatics wins on language coverage, on-premises deployment, and compliance certification depth. It also pairs naturally with OpenAI Whisper benchmarks as a reference point: Whisper is open-source and strong on accuracy, but offers no enterprise SLA, no diarization, and no streaming, making Speechmatics the structured alternative for teams that have outgrown self-hosted Whisper.

The compliance and deployment reality

Speechmatics' clearest differentiation from both Deepgram and AssemblyAI is its deployment flexibility, and the compliance certification stack that backs it up. The on-premises option is not a marketing checkbox. Flow On-Premise, released in 2024, deploys a full conversational AI stack (STT, LLM turn-taking, and TTS) inside the customer's own infrastructure via Docker containers or preconfigured virtual appliances. This means audio never leaves the customer's firewall, which matters enormously in healthcare (HIPAA), European government (GDPR with stricter national data residency rules), defense (classified audio), and financial services (MiFID II call recording requirements).

Certifications held as of April 2026: ISO/IEC 27001:2022, SOC 2 Type II, GDPR compliance, HIPAA compliance. Data is encrypted at rest with AES-256 and in transit with TLS 1.2 minimum. The infrastructure runs on Kubernetes-controlled containers with SIEM monitoring. Enterprise customers can also select multi-region cloud configurations to satisfy national data residency requirements without going fully on-premises.

For teams evaluating Speechmatics against ElevenLabs for combined STT plus TTS workflows: ElevenLabs is primarily a TTS product with recent STT additions, while Speechmatics has 20 years of STT specialization with TTS added more recently (the free tier includes 1M characters/month of English TTS). The product postures are inverted. For voice agents that need best-in-class STT with the option to swap in a specialist TTS layer, Speechmatics pairs cleanly with ElevenLabs or Coqui TTS via the Flow API architecture rather than competing directly. Teams building fully hosted voice pipelines without compliance constraints might find Coqui TTS or ElevenLabs more feature-complete on the voice synthesis side.

Where Speechmatics reliably falls short

The recurring frustration threads across G2, Capterra, and Gartner Peer Insights land in four areas:

Pricing opacity at scale. The Pro tier has published rates and a 6,000 hours/month cap, but anything above that requires an Enterprise conversation with no public pricing. Users in review threads specifically flag that pricing structure becomes complex when scaling up. The volume discount (20% above 500 hrs/month, additional breaks from 24,000 hrs/year) helps, but you cannot build a reliable cost model for large-scale production without speaking to sales.

Custom Dictionary limitations. The JSON-based Custom Dictionary is useful for domain-specific terminology but caps entries at 6 words. Healthcare and legal users needing to specify long compound clinical terms or multi-word legal phrases find this restrictive. AssemblyAI handles this via natural language prompting without a word cap.

Streaming latency configuration overhead. The `max_delay` parameter defaults to 4 seconds, which is fine for post-processing use cases but creates friction for developers building real-time voice agents. The 0.7-second minimum is achievable, but requires documentation reading and testing that Deepgram's voice agent products avoid by design.

Limited built-in post-processing. Speechmatics returns accurate transcripts. It does not return summaries, sentiment labels, entity extractions, or structured call analytics out of the box. Every feature beyond raw transcription requires a separate pipeline and integration. Teams expecting the kind of turnkey analytics that AssemblyAI's LeMUR provides will need to build that layer themselves.

Who Speechmatics is for

Build with Speechmatics when: you need on-premises or private cloud deployment and cannot accept audio leaving your infrastructure; your audio spans more than 6 languages in streaming mode; you serve a regulated industry (healthcare, finance, defense, government) where HIPAA/SOC 2/ISO 27001 certification is a procurement requirement; you need robust speaker diarization included without add-on pricing; your live captioning or broadcast use case needs sub-second latency across dozens of languages simultaneously.

Skip Speechmatics when: you are a solo developer or small startup exploring STT for the first time (Deepgram's developer experience and pricing are more accessible at low volumes); you need built-in LLM summarization, sentiment analysis, or entity extraction in a single API call (AssemblyAI LeMUR handles this without a separate pipeline); you are building a US-based voice agent where latency is the primary variable and compliance is not a gating factor (Deepgram Flux is purpose-built for this); or your budget requires a genuinely unlimited free tier before committing to usage-based pricing (the 8 hours/month free tier is a starting point, not a production environment).

The Limecraft benchmark story is illustrative. The media collaboration platform evaluated multiple ASR vendors head-to-head before standardizing on Speechmatics for its broadcaster customers. Maarten Verwaest, Limecraft's CRO, reported: "They consistently outperform other vendors for word error rate and punctuation, playing a pivotal role in the development of our workspace." That is the kind of buyer Speechmatics is built for: enterprise teams that have run the benchmark, compared the compliance docs, and found that accuracy plus deployment flexibility outweighs the premium cost.