Deepgram

Deepgram is the long-running speech AI platform that has quietly become the default STT engine behind a large share of production voice agents, contact-centre analytics tools and meeting bots. The Nova-3 STT model delivers state-of-the-art word error rate across 30+ languages with sub-300ms streaming latency, includes diarisation, smart formatting and keyword boosting, and runs cheaper-per-minute than competing managed providers. Deepgram also ships Aura, a streaming TTS model designed for low-latency voice agents, and the Deepgram Voice Agent API, a single endpoint that combines STT, an LLM of your choice and Aura TTS with turn-taking handled server-side — the cleanest way to ship a phone-able agent if you want one vendor end-to-end. Beyond real-time, Deepgram has strong batch transcription for podcast and video workflows with topic detection, entity extraction, summarisation and translation. New customers start with a \$200 credit, then pay metered per-minute rates that scale down with volume, and enterprise customers can run Deepgram fully on-prem for HIPAA and air-gapped use cases. Deepgram remains the default choice when accuracy per dollar matters more than brand cachet.