Vibeleaderboard
Deepgram

ai lab

Deepgram

Deepgram matters because it develops and operates both the listening and speaking layers needed by production voice agents. Its strongest contribution is not a single benchmark headline but the integration of streaming recognition, domain adaptation, conversational timing, synthesis, and deployment controls in one specialist platform. Independent results also make it a useful reminder that speech quality must be judged on the actual audio, latency, language, and cost constraints of a use case.13,12,2,3

Official site ↗Models in the Atlas →

Profile

Overview

From physics signals to speech infrastructure

Deepgram is a speech AI company founded in 2015 by Scott Stephenson and colleagues with backgrounds in experimental particle physics and machine learning. The company's account traces the original work to neural analysis of signals from a dark-matter detector, followed by end-to-end speech recognition research at the University of Michigan. It joined Y Combinator in 2016 and developed a developer API around models it trains and serves rather than acting only as an interface to outside speech vendors.7,8,9

Nova and Aura become a broader voice stack

The product expanded from transcription into a speech stack. Nova is the automatic speech-recognition family, Aura covers generated speech, and the Voice Agent API coordinates recognition, language-model calls, and synthesis. Nova-3 added multilingual transcription, code switching, diarization, and keyterm prompting. Live evaluations place Nova-3 competitively on some workloads, while independent ASR research shows why word-error rankings change substantially with dataset, streaming mode, language, and normalization.10,13,2,3,6

Flux models conversational timing

Flux introduced a different boundary between transcription and dialogue timing. Deepgram describes Flux STT as a single model that jointly produces transcripts and estimates conversational end-of-turn state rather than combining a conventional recognizer with a separate voice-activity detector. That design is aimed at interruptions and response timing in voice agents. The claim is technically specific, but published comparisons are mainly Deepgram's internal benchmarks and should not be treated as independent proof of universal leadership.11,12

Scale grows, and comparison stays workload-specific

Deepgram entered 2026 with greater financial and operational scale. TechCrunch reported a $130 million Series C at a $1.3 billion valuation and the acquisition of restaurant voice-automation company OfOne. The company has continued releasing speech products while independent benchmarks expose a real tradeoff between latency, accuracy, language, and price. That evidence supports the strategy of a vertically integrated voice-model provider without resolving which vendor is best for every audio domain.1,2,4

Notable contributions

  1. 01Commercial end-to-end neural speech specializationDeepgram built its early platform around end-to-end learned speech recognition and direct neural processing of audio. This was an early commercial specialization, not the invention of end-to-end speech recognition itself.7,8
  2. 02Keyterm prompting in Nova-3Nova-3 lets developers bias recognition toward domain-specific terms at inference time without retraining a custom model, making vocabulary adaptation part of the hosted API workflow.10,13
  3. 03Turn-aware transcription for voice-agent workflowsFlux STT jointly models transcript content and conversational state so a voice agent can decide when to respond without relying only on a separate silence or activity detector. The priority claim is limited to Deepgram's documented production model design.11,12