Deepgram

Deepgram is a hosted API for speech, not an open-source library: a family of proprietary ASR models (Nova-3, and the newer Flux model built for real-time use), a text-to-speech line (Aura-1/Aura-2), and a Voice Agent API that chains STT, an LLM (Deepgram's own or a bring-your-own model), and TTS into one WebSocket connection. The client SDKs (Python, JS/TS, .NET, Go) are open source and actively maintained — the Python SDK shipped a new PyPI release on 2026-08-12, and both the Python and JS SDK repos on GitHub had commits within the last 24-48 hours as of this check — but the models themselves are closed and run only on Deepgram's infrastructure or, for enterprise customers, in self-hosted containers on customer-owned NVIDIA GPUs.

On Artificial Analysis's streaming speech-to-text leaderboard (read directly off the rendered page on 2026-08-19, not vendor copy), Deepgram Flux has the lowest time-to-final-transcript of any model tracked — 0.02 seconds from detected end-of-speech to final transcript, more than 2x faster than the next fastest (Soniox v5 Real-Time at 0.05s) and roughly 20x faster than several other proprietary competitors in the same comparison. For a voice agent, that latency is often the difference between a conversation that feels natural and one that feels like a phone tree, so this is a real and specific technical edge, not a marketing number.

On the same independent benchmark's non-streaming Word Error Rate index (AA-WER v2, a duration-weighted average across three real-world datasets), Deepgram's Nova-3 scored 5.2% WER, ranking 29th of the 31 models shown (out of 63 tracked total) as of 2026-08-19 — behind ElevenLabs Scribe v2 (2.2%), AssemblyAI's Universal-3.5 Pro (3%), and OpenAI's Whisper Large v3 (4.1%), among others. This is close to Deepgram's own published claim of a ~5.26% median WER, so the vendor isn't overstating its own number, but a buyer choosing purely on transcription accuracy has several models that currently score meaningfully better. Deepgram's real differentiator, per the same source, is speed rather than peak accuracy.

Pricing is fully public and unusually granular: pay-as-you-go rates are quoted per minute (or per hour equivalent) down to the tenth of a cent, for every model and feature combination (verified against deepgram.com/pricing on 2026-08-19). Representative pay-as-you-go rates: Nova-3 streaming transcription $0.0048/min ($0.29/hr) monolingual, $0.0058/min multilingual; Nova-3 batch/pre-recorded $0.0043/min monolingual; Flux (real-time) streaming $0.0065/min; Aura-2 text-to-speech $0.030/min; the Voice Agent API (Standard tier) $0.056/min end-to-end. Every new account gets $200 in free credit with no credit card required. A Growth plan requires a prepaid commitment starting at $4,000/year in exchange for roughly a 12.5-20% discount across products, higher concurrency limits, and priority support; add-on features (redaction, keyterm prompting, sentiment/intent/topic detection) are billed separately, on top of the base transcription rate, which is easy to miss when estimating a per-minute cost from the headline number alone.

Deepgram publishes SOC 2 Type 1 and Type 2 certification and states it is HIPAA compliant, signing Business Associate Agreements for eligible customers (deepgram.com/pricing FAQ, verified 2026-08-19). For data-sovereignty or regulatory needs, it offers self-hosted containers deployable in a customer's own VPC or on-premises, which requires NVIDIA GPUs and is positioned as an Enterprise-tier offering rather than something available on usage-based plans.

Deepgram raised a $130M Series C in January 2026 at a $1.3B valuation, led by AVP, with participation from existing investors including Alkeon, In-Q-Tel, Madrona, Tiger, Wing, Y Combinator, and BlackRock-managed funds, plus new strategic investors Twilio, ServiceNow Ventures, SAP, and Citi Ventures (deepgram.com press release and Yahoo Finance coverage, verified 2026-08-19). Total disclosed funding is reported at roughly $234M across 12 rounds by Tracxn; that aggregator figure is third-party and not independently verified here. Employee headcount estimates vary widely across data providers (from roughly 180 to 340) with no primary source giving a single authoritative number, so headcount is not established.

G2 and Trustpilot both blocked automated access (403/CAPTCHA) during this review, including via a headless browser, so no G2 or Trustpilot rating or review count could be verified from a primary source as of 2026-08-19; any specific star rating or review count attributed to Deepgram found via search aggregation should be treated as unverified.