Speechmatics Speech-to-Text

Speechmatics Speech-to-Text is a neural-network-based transcription service founded in 2006 and headquartered in Cambridge, United Kingdom. It serves enterprises and developers who need high-accuracy, multilingual transcription for voice agents, live captioning, meeting analytics, and real-time translation. The platform is designed for users who prioritize accuracy and customization over raw speed, and it supports both real-time and batch processing across over 55 languages and dialects. Its target audience includes media companies like AI Media (which scaled live captioning 120x using Speechmatics), technology partners like LiveKit (enabling 100,000+ developers with speech recognition), and enterprises requiring on-prem, cloud, or hybrid deployment for data security or low-latency requirements.

The system uses end-to-end neural networks with contextual biasing to improve recognition of domain-specific terms, and its Ursa models allow configurable latency-accuracy trade-offs via a max_delay parameter. Key capabilities include speaker identification (diarization), multi-channel audio support for separating overlapping speakers, and customizable transcription with custom vocabularies and formatting rules. In a notable achievement, Speechmatics eliminated 40% of transcription errors in Norwegian speech recognition, reducing mistakes by 1,400 per hour of audio. It also offers text-to-speech alongside STT, with pricing that varies from free tiers to enterprise plans. The platform is named best-in-class STT for voice agents based on benchmark results, and it powers use cases like virtual meeting analytics and real-time translation for clients such as Adobe (on-device for Premiere) and NCI (99% increase in automated captioning usage).

Speechmatics competes against hyperscalers like Google Cloud Speech-to-Text, Amazon Transcribe, and Microsoft Azure Speech, as well as specialized providers like AssemblyAI and Deepgram. While hyperscalers offer broad language coverage and moderate pricing, Speechmatics differentiates with higher accuracy in noisy or accented speech, flexible deployment (on-prem, cloud, hybrid), and specialized language support like Norwegian and Persian. In independent benchmarks, Speechmatics’ Ursa models demonstrated configurable latency/accuracy trade-offs, but its latency is not the lowest in the market compared to Deepgram’s real-time streaming. Its pricing structure is more complex than competitors like AssemblyAI, which offers simpler per-minute rates, and its enterprise plans require direct sales engagement for custom quotes.

The honest trade-offs: Speechmatics’ accuracy is industry-leading, but its latency is higher than some real-time-focused alternatives, and accuracy can decrease at scale with very long audio files. Its pricing is higher than many competitors, especially for small-scale usage, and the structure is complex with multiple tiers and add-ons. On-prem deployment requires significant infrastructure investment, and the free plan limits concurrent sessions to 2. However, for organizations that need high accuracy across 55+ languages, customizable vocabularies, and flexible deployment options (including air-gapped environments), Speechmatics provides a robust solution that justifies the premium cost.

Free plan: 2,400 minutes/month STT with 2 concurrent sessions, 1M characters TTS. Pro plan: same minutes with 50 concurrent sessions, 20% discount on overage. Enterprise plan: unlimited STT and TTS, customization, support, security, and on-prem/cloud/hybrid deployment (custom pricing).