Deepgram

Deepgram built its reputation on a simple claim: transcribe speech faster and cheaper than anyone at production scale. That focus made it the plumbing behind a large share of the voice-agent boom — when a phone bot answers you with no awkward pause, there's a decent chance Deepgram is doing the listening.

What we like

Speed and price are the headline. Nova models stream transcripts with latency low enough for natural conversation, and per-minute costs undercut most rivals — at call-center volumes the difference is a line item someone notices. The $200 starting credit is the most generous evaluation runway in the category, enough to build and stress-test a real prototype. The Voice Agent API is a shrewd move up the stack, bundling STT, LLM orchestration, and its Aura TTS voices into one socket so you're not conducting a three-vendor latency orchestra. On-prem deployment wins deals in healthcare and finance that cloud-only competitors can't touch.

Where it falls short

Deepgram is a components company: there's no app, no UI for a podcaster or journalist. On heavily accented or noisy audio, some benchmarks show newer multilingual models edging it on raw accuracy, so test your actual audio rather than trusting anyone's marketing page. Aura's TTS voices prioritize latency over expressiveness — fine for a support bot, flat for narration.

Verdict

For real-time voice applications — agents, IVR replacement, live captioning — Deepgram is the default we'd start with, and the free credit makes that start free. For transcript analytics and understanding, compare AssemblyAI. If you're not building software, look elsewhere entirely.