Deepgram

Deepgram is a voice-AI infrastructure company whose Nova speech-to-text models are known for being both fast and cheap enough to power real-time products — a common choice for contact centers, meeting notetakers, and voice assistants. Nova-3 handles more than 30 languages with word-level timestamps, diarization, keyword boosting, redaction, and topic detection, and streams transcripts in real time with very low latency. Deepgram's Aura text-to-speech engine produces natural voices with sub-300ms first-audio latency, which makes it viable for voice agents where round-trip time to the user matters. On top of those two, Deepgram now ships a Voice Agent API that stitches STT, an LLM of your choice, and TTS into a single WebSocket connection — a much simpler integration path if you're building a Twilio- or WebRTC-based voice assistant. Pricing is pay-as-you-go with a $200 free credit for new accounts: pre-recorded Nova transcription is around $0.0043/minute, live streaming is around $0.0077/minute, and Aura TTS is priced per character. Larger customers get committed-use discounts and a self-hosted option. Deepgram is a good fit for engineering teams that need voice at real-time speed and don't want to pay per-hour rates that scale poorly with millions of minutes.

What Deepgram does in practice

Nova speech-to-text with 30+ languages and word-level timestamps. Real-time streaming transcription with very low latency. Aura text-to-speech with sub-300ms first audio. Voice Agent API bundling STT + LLM + TTS over one WebSocket. Speaker diarization, PII redaction, keyword boosting, and topic detection. Self-hosted deployment for regulated industries. These capabilities should be judged as a workflow, not isolated checkboxes. Concrete best-fit scenarios include: Real-time voice assistants for support, sales, and IVR. Contact-center transcription and analytics at scale. Live captioning and accessibility tooling. Meeting notetakers and voice-first productivity apps.

Pricing and total cost

Free trial is listed at $200 credit; New accounts get $200 in free API credit.. Pay-as-you-go — Nova pre-recorded is listed at ~$0.0043/min; Batch transcription at high accuracy with per-minute billing.. Pay-as-you-go — Nova streaming is listed at ~$0.0077/min; Real-time WebSocket streaming transcription.. Aura TTS is listed at Per-character; Natural voice synthesis, priced per character with volume tiers.. Enterprise / self-hosted is listed at Custom; On-prem deployment, committed-use discounts, dedicated support.. These figures come from the staging record and vendor pages were fetched during this revision, but pricing is mutable. Recheck the live pricing page, checkout, or quote before publishing or buying. Include required seats, usage credits, model or API charges, storage, implementation, support, taxes, and overages. A free tier can prove compatibility; it does not necessarily model production volume or support needs.

Strengths and tradeoffs

Practical strengths include: Nova speech-to-text with 30+ languages and word-level timestamps. This directly supports the primary workflow. Real-time streaming transcription with very low latency. Aura text-to-speech with sub-300ms first audio. Voice Agent API bundling STT + LLM + TTS over one WebSocket. Important constraints include: At least one tier requires a sales quote, preventing a complete public budget calculation. Features and limits can change by plan, billing interval, region, and release; verify current documentation. No native MCP support is documented in this record, so agent workflows may require API work or another integration layer. A representative pilot is necessary because a feature list does not establish reliability or operational fit. A capable product can still be the wrong operational choice. Document who reviews outputs, who owns failures, what data may enter the system, and how work continues during an outage.

How to evaluate Deepgram

Run a controlled media test before buying. Use at least 20 recordings representing quiet speech, background noise, overlapping speakers, accents, product names, and poor microphones. Measure word error rate on a human-corrected sample, speaker-label accuracy, time to first result, correction minutes per recorded hour, and total cost per usable hour. For live workloads, record p50 and p95 latency and test reconnect behavior. Confirm consent, recording notices, retention, deletion, regional storage, and whether model training is enabled. A polished demo on clean audio is not evidence that the system will handle real calls.

Bottom line

Deepgram is worth shortlisting when its capabilities match a measured workflow and the buyer can verify current limits. Do not select it from a feature list alone. Compare at least two alternatives using the same inputs, acceptance criteria, and cost model. Preserve test results so renewal is based on outcomes rather than adoption anecdotes.