Tool Introduction

Dr. Tony Robinson, who pioneered applying recurrent neural networks to speech recognition, founded Speechmatics in Cambridge, England in 2006; the business traded as Cantab Research Ltd before rebranding.

Public milestones include a billion-word text corpus released into the public domain in 2014, the accent-independent Global English language pack in 2018, Global Spanish in 2020, Autonomous Speech Recognition in 2021, and the Flow voice-agent API in 2024. A £6.35 million Series A closed in 2019 with Albion Venture Capital, IQ Capital, and Amadeus Capital Partners, alongside the Queen's Award for Enterprise in Innovation.

Demo Available?

Yes. The Free tier starts instantly with $100 credit and no card, Pro offers Try Pro for free, and Enterprise buyers book a demo.

Tool's Accuracy

Speechmatics measures transcription quality as word error rate (WER), the sum of insertions, deletions, and substitutions divided by the reference word count, computed with Levenshtein distance on test sets of at least 10,000 words per language.

Independent benchmark: a Pipecat benchmark dated August 2026 tested 12 speech-to-text and voice-AI services and recorded a 1.07% pooled word error rate for Speechmatics, the lowest of the 12.
Model ranking: documentation describes Enhanced as the highest accuracy option, Standard as faster with slightly lower accuracy, and the early-access multilingual Melia 1 as matching Standard for accuracy.
Medical model: Speechmatics states its Medical Model reduces errors on key terms by up to 50%, a self-reported claim with no disclosed method.

Aggregate Rating

4.7/5 stars (based on 72+ verified reviews across the internet)

Detailed Overview

What Speechmatics Does

Speechmatics turns spoken audio into text that holds up on messy recordings: accented speech, several people talking at once, background noise, and speakers switching languages mid-sentence. Teams either send finished files to the Batch API or stream live audio over a WebSocket connection to the Realtime API. Transcripts come back with speakers separated and with domain terms recognized through a Custom Dictionary of up to 1,000 entries.

Who Speechmatics Is For

Engineering and product teams embedding speech into their own applications are the core audience, not consumers shopping for a transcription app. A solo developer can prototype straight against the API, while contact centers, court reporters, broadcasters, and clinical-intake vendors arrive with harder problems: noisy phone calls, live television that cannot wait, or patients speaking any of dozens of languages.

Top Differentiators

An independent Pipecat benchmark of 12 speech-to-text services measured Speechmatics at a 1.07% pooled word error rate, the lowest in that test. Deployment flexibility is the second point of difference, since the same engine runs in multi-region cloud, Docker, Kubernetes, an on-premises Virtual Appliance, or on-device, which matters where audio cannot leave the building. Speaker locking holds one target voice and filters out background talkers.

Founding Year

2006

Target Industry

B2B, Enterprise, SMB, Developers, Healthcare, MedTech, Media and Broadcast, Legal and Court Reporting, Contact Centers, Education, Food Service

Available Integrations

Vapi: Speechmatics is selectable as a transcription provider inside Vapi's hosted no-code voice-agent builder, and is the only transcriber there providing speaker diarization.
LiveKit Agents: the official plugin supplies both speech-to-text and text-to-speech in a LiveKit Agents pipeline; the guide assumes LiveKit Cloud, and self-hosting needs extra manual setup.
Pipecat: open-source, code-first voice pipeline using both speech-to-text and text-to-speech, testable locally at localhost:7860/client, with a longer first run while models load.
Zapier: labeled in beta, with two actions, Get Transcript from Audio for short recordings and Submit Audio (Webhook) for longer recordings returned asynchronously.
Telnyx: a named hosted-platform partner offering Speechmatics recognition as an embedded capability, with no dedicated integration guide published.
Voice SDK: a Python software development kit (SDK) for building responsive voice agents directly against the API.
Realtime SDKs: Python and JavaScript kits for streaming transcription inside custom pipelines, with no stated beta flag or plan restriction.
Batch SDKs: Python and JavaScript kits for submitting pre-recorded files programmatically.
.NET SDK: a Realtime kit for teams building transcription into .NET applications.
Rust SDK: a combined Realtime and Batch kit for Rust codebases.
Text-to-Speech SDK: a dedicated Python kit for speech synthesis requests.

Free: a $100 free credit grant with no card required, 2 concurrent real-time sessions, 55+ languages for Speech-to-Text plus Text-to-Speech in English; accounts created before August 1, 2026 also receive a one-time $25 transition credit.
Pro: pay-as-you-go at $0.129 per unit with no commitment required, 50 concurrent real-time sessions, 10 file jobs per second across 55+ languages, and online email support.
Enterprise: contract-based custom pricing with no rate limits, custom models and voices, privacy-first on-prem and on-device deployment options not available on other tiers, dedicated customer support, onboarding, and solutions engineering, engaged through Book a demo or Speak to sales.

Pro requires a payment card on the Billing page to activate pay-as-you-go; granted and free credits are drawn down first, and only overflow usage is charged to the card. Billing occurs on the 1st of each month for the previous month's usage, calculated to the exact second based on the per-hour rate. A 20% discount applies automatically above 500 hours of usage per product type per month, with additional discounts available starting from 24,000 hours per year, and an opt-in model-training program offers 33% off Speech-to-Text rates.

Pricing last verified: 09/09/2026
Disclaimer: Pricing is subject to change. Please confirm current pricing on the vendor site.

Features & Specs

Language coverage: 55+ languages, with the Melia 1 model handling code-switching across 56 languages in Realtime; named examples include Arabic, Hebrew, Persian, Bengali, Hindi, Tamil, Mandarin, Cantonese, Japanese, Korean, Swahili, and Welsh.
APIs: a Batch API for file-based transcription, a WebSocket-based Realtime API for streaming, and a Management API for projects and API keys.
Models: Enhanced, Standard, and Melia 1.
Stated accuracy: 90%+ accuracy in the real-world, 96.9% sequence accuracy on character strings, 98.0% on digits, and 85.4% on mixed alphanumerics.
Latency: described as sub-500ms.
Custom Dictionary: up to 1,000 domain-specific terms.
Speaker diarization: speaker separation available in both batch and realtime jobs.
Speech Intelligence: batch-only sentiment analysis, summarization, topic detection, and auto-chapters.
SDKs: Python, React, and JavaScript software development kits (SDKs).
Deployment: multi-region cloud, Docker on CPU or GPU, Kubernetes, an on-premises Virtual Appliance, and on-device.

Security & Compliance

Speechmatics holds ISO/IEC 27001:2022 accreditation and SOC 2 Type II certification, and states it is fully compliant with the Health Insurance Portability and Accountability Act (HIPAA) and with GDPR. Data is encrypted with AES-256 at rest and TLS 1.2 or higher in transit, and the company states it will never share customer data with any unauthorized third party.

Cloud hosting: Microsoft Azure with built-in security services and a SIEM (Security Incident and Event Monitoring) solution.
On-premises infrastructure: Kubernetes-controlled containers.
Testing: regular penetration testing by CREST-accredited providers.
Partnership: a stated security partnership with Kroll.
Model training: an opt-in program that gives 33% off Speech-to-Text rates to customers who join it.

Industry Use-Cases

Multilingual clinical intake

Humetrix builds multilingual clinical-intake software for healthcare providers covering 27 languages, and it moved off OpenAI's Whisper v3 large model after accuracy on drug names, anatomical terms, and condition names proved inadequate for clinical safety. Speechmatics runs on-premises inside AWS elastic containers for that deployment. Supporting healthcare services at the Paris 2024 Olympics, clinical staff triaged a patient population spanning 97 countries and handled 22+ languages, identifying medications whatever the language spoken. Details here come from the company's published case studies.

Live television captioning

AI-Media, a captioning and media-accessibility company, embedded the Speechmatics transcription engine in LEXI 3.0, its live-captioning service delivered over the LEXI Network cloud platform. Volume was the constraint that mattered. The company delivers 120 times more content in 2025 than five years earlier, a company-wide scaling metric rather than an effect isolated to this integration, and its chief executive describes the engine as really optimized for live transcription with consistent quality across global media markets.

Cross-border media monitoring

Roughly 2 million print pages and 3,000+ broadcast channels across six continents in 20+ languages pass through Media Track every month, and the company tested multiple engines before integrating Speechmatics in 2023. Head-to-head word error rates against Google, measured by Media Track itself, came in at 5.6% versus 24.5% for English, 23.1% versus 48.4% for German, and 33.1% versus 63.9% for Swedish. Speech-to-text accuracy improved 20% overall between Q4 2023 and Q1 2024.

Tool's Alternatives

Deepgram serves the same developers and enterprises with streaming and batch speech-to-text plus diarization, and positions itself as conversation-aware, natively handling turn-taking, interruptions, and conversational context inside one platform that also bundles text-to-speech and LLM (large language model) orchestration behind a single API. Its homepage does not surface accuracy, language-coverage, or technical-limit figures for the core speech-to-text product, so those specifics require digging through documentation.

Google Cloud Speech-to-Text suits enterprise buyers already inside Google Cloud, covering short-audio, long-audio batch, and streaming recognition with a model lineup of Chirp 3, Chirp 2, and older models, plus WebVTT and SRT caption export. Native ties to IAM, storage, and other Google AI APIs set it apart. Custom speech models and on-premises or edge deployment are allowlist-only features that need special Google approval.

Amazon Transcribe fits teams building call-center, media, and general transcription workloads on AWS, with streaming and batch jobs, diarization, custom vocabulary, and language detection across 100+ languages. Specialized tiers distinguish it: Transcribe Call Analytics for sentiment and summary extraction, and a separate HIPAA-eligible Transcribe Medical product for clinical documentation. Its product page states no explicit accuracy, latency, or file-size limits.

Microsoft Azure AI Speech handles call-center and meeting transcription for enterprise developers across 100+ languages, and lets buyers choose between Microsoft's own models and a hosted latest OpenAI Whisper model, with container-based embedded deployment for intermittent connectivity. Language coverage is described only as an ever-growing set rather than an exact count, and usage-hours billing replaces transparent tiers on that page.

Gladia is a smaller API-first challenger, EU-headquartered with EU data residency by default, claiming 100+ languages, native mid-sentence code-switching, sub-300ms real-time latency, and diarization included rather than charged as an add-on, alongside SOC 2 Type II, GDPR, HIPAA, and ISO 27001. No on-premises or self-hosted deployment option appears on its site.

Frequently Asked Questions

What kind of speech tasks does Speechmatics handle?

Speechmatics converts speech to text in batch and real time, generates English text-to-speech, returns translation alongside transcription from a single audio submission, and powers turn-based voice agents through the Flow API.

How many languages does the platform support?

Speech-to-text covers 55+ languages, and the Melia 1 model handles code-switching across 56 languages in Realtime. Roughly 30+ of the supported languages have active translation pairs. Text-to-Speech is low-latency English.

What does Speechmatics cost to start using?

New accounts receive $100 in free credit with no card required. Pro is pay-as-you-go at $0.129 per unit with no commitment. Enterprise pricing is contract-based and custom, arranged through sales.

Can Speechmatics run on-premises instead of the cloud?

Yes. Deployment spans multi-region cloud, Docker containers on CPU or GPU, Kubernetes, an on-premises Virtual Appliance, and on-device. Privacy-first on-prem and on-device options sit in the Enterprise tier.

Which certifications and compliance standards does Speechmatics hold?

Speechmatics holds ISO/IEC 27001:2022 accreditation and SOC 2 Type II certification, and states HIPAA and GDPR compliance. Encryption is AES-256 at rest and TLS 1.2 or higher in transit.

How accurate is Speechmatics against competing services?

An independent Pipecat benchmark dated August 2026 tested 12 services and measured a 1.07% pooled word error rate for Speechmatics, the lowest of the group. Enhanced ranks above Standard in documentation.

Are volume discounts available on transcription usage?

A 20% discount applies automatically above 500 hours per product type each month, with further discounts from 24,000 hours per year. Opting into model training takes 33% off Speech-to-Text rates.

Which voice agent frameworks integrate with Speechmatics?

Official guides cover Vapi, LiveKit Agents, and Pipecat, with Zapier offered in beta for no-code automation. First-party kits ship for Python, React, JavaScript, .NET, and Rust.

How does Speechmatics handle noisy multi-speaker audio?

Speaker diarization separates voices in batch and realtime jobs, while speaker locking isolates one target speaker and filters background talkers for voice agents. A Custom Dictionary adds up to 1,000 domain terms.

Comments are closed.

Speechmatics

Speech APIs powering Voice AI