Deepgram

Deepgram is a developer-first speech AI platform whose core product is a low-latency, high-throughput speech-to-text (STT) API. Its flagship Nova-3 model targets real-time, production-scale voice applications, contact centers, voice agents, meeting/transcription tooling, and media workflows, where streaming latency and cost-per-hour matter as much as raw accuracy. Deepgram differentiates on speed (sub-300ms streaming latency), aggressive per-minute pricing, and a clean API surface with official SDKs across the major backend languages. Beyond STT it now ships a Text-to-Speech line (Aura-1/Aura-2) and a Voice Agent API, positioning itself as a full voice stack rather than a transcription-only vendor.

Deepgram's strongest fit is real-time, high-volume English (and increasingly multilingual) workloads where its streaming latency and price-per-hour beat incumbents like Google, AWS Transcribe, and Whisper-based services. Its own benchmarks claim a 5.26% batch WER and a 54.2% streaming WER reduction versus the next-best competitor, but those are vendor-run on Deepgram-selected datasets; independent third-party measurement (Artificial Analysis) puts real-world Nova-3 WER closer to ~18%, so buyers should validate on their own audio. Nova-3 added keyterm prompting (up to 100 terms / 500 multilingual tokens) to recover domain accuracy, plus real-time multilingual code-switching and 30+ language support.

The main trade-offs are accuracy on hard inputs (strong accents, overlapping speakers, less-common languages), cost at very large volumes, and some model-maturity gripes (users report Nova-2 occasionally outperforming Nova-3, and feature gaps between model generations). For teams that need maximum accuracy on noisy multi-speaker audio over raw speed, AssemblyAI or Whisper-class models are common alternatives. But for latency-sensitive voice agents and streaming transcription at scale, Deepgram remains a top, frequently top-rated, choice on G2 (4.6/5).