Speech APIs powering Voice AI
Grade: B — Score: 80/100
Speechmatics provides speech APIs for speech-to-text and text-to-speech, with live transcription, multilingual support across 55+ languages, speaker diarization, custom vocabulary, and deployment options on device, on prem, and in the cloud.
The workflow is to create an API key, connect your workflow, and use the documentation, SDKs, and native integrations to build use cases like voice agents, live captioning, meeting notes, contact center analytics, and medical dictation.
Risks and constraints are not fully stated on the site; the company says it does not log data as standard, and some plans are usage-based or custom-priced depending on scale and deployment needs.
Free: $0
Pro: Usage-based, from $0.12 per hour
Enterprise: Custom pricing
Buyers usually choose Speechmatics for its 55+ language coverage in real-time streaming, flexible deployment including on-premise and on-device, and voice-agent building blocks like turn detection. Deepgram tends to attract teams optimizing for streaming cost efficiency, while AssemblyAI appeals to teams wanting transcription plus built-in speech intelligence in one API. Test all three on your own audio, since accuracy varies with language and conditions.
Speechmatics bills by usage in credits, where one credit equals one US dollar. Batch transcription starts at $0.12 per hour for the Melia 1 model, with Standard at $0.23, Enhanced at $0.38, and the medical model at $0.15 per hour. Text to speech costs $0.011 per 1,000 characters with the first 1 million free. Enterprise plans use custom contract pricing.
Yes. New accounts receive a $100 credit grant at no cost with no payment card required, and the Free plan lets you use those credits before adding a card to move to Pro. The vendor's own homepage confirms you can test transcription quality and language support this way.
Speechmatics supports 55+ languages, covering accents, dialects, and code-switching where speakers move between languages mid-sentence. Its Melia 1 model transcribes all of them in a single pass without pre-selecting a language.
The vendor states compliance with ISO 27001, GDPR, HIPAA, and SOC 2 Type II, and its trust center documents additional frameworks. Combined with on-premise and on-device deployment options on Enterprise plans, this makes it a common choice for healthcare, legal, and financial services workloads.
Yes. Speaker diarization identifies and separates individual speakers in meetings, calls, legal proceedings, and healthcare consultations. Voice agent builds can also use speaker focus to concentrate on or ignore specific speakers.
In an independent August 2026 benchmark by Pipecat cited on the vendor's homepage, Speechmatics returned a 1.07% pooled word error rate, the lowest of the 12 services tested. Real-world accuracy still depends on your audio, languages, and conditions, so the vendor recommends testing with your own recordings.
Yes, but only on Enterprise contract plans, which include on-premise and on-device deployment so audio never leaves your infrastructure. Self-serve Free and Pro plans run in the vendor's cloud.
Only if you opt in. Model training, also called data logging, is a workspace setting that lets Speechmatics use anonymized data to improve its models in exchange for a 33% discount. You can change the setting at any time and it applies to future usage only.
Speechmatics positions itself as the multilingual and deployment-flexible specialist, with 55+ real-time languages, on-premise and on-device options, and no cloud lock-in. Google Cloud Speech-to-Text ties your pipeline to Google Cloud and charges more for real-time. The right pick depends on your language needs, deployment constraints, and existing cloud commitments.