Deepgram

What it is

Deepgram is the speech layer many voice-support stacks run on without their buyers knowing it. It sells speech-to-text (the Nova-3 family for transcription and Flux for real-time conversation), text-to-speech (Aura-2 and the newer Flux TTS), and a Voice Agent API that packages transcription, an LLM, and a voice into one WebSocket at a flat hourly rate. When you configure a Vapi assistant or a Genesys virtual agent, Deepgram is often the transcriber sitting underneath it.

This page answers one question: which speech layer sits under your support voice agent, and should you buy Deepgram directly or take whatever your voice platform bundles? The company is independent and well funded. It raised a $130M Series C at a $1.3B valuation on 13 January 2026, led by AVP, with Twilio, ServiceNow Ventures, SAP, and Citi Ventures joining as strategic investors. In the same announcement it acquired OfOne, a drive-thru voice AI company, which now anchors its restaurant offering.

Why it shows up in CX and support stacks

• Flux was built for turn-taking, not transcripts. Launched in October 2025, Flux does transcription and end-of-turn detection in one model instead of bolting a voice-activity detector onto an ASR stream. Deepgram cites about 260ms end-of-turn latency and ships native barge-in and “eager” end-of-turn events so your LLM can start drafting before the caller finishes. Those are vendor numbers. What you can check yourself is the effect: fewer cases of the agent talking over the caller or sitting in dead air.
• Domain models for the calls that break generic STT. Nova-3 Medical gained ten-language support on 6 October 2026, and Nova-3 Pharma (17 September 2026) targets drug-name recognition. Keyterm prompting lets you bias recognition toward SKUs, plan names, and account IDs, and since 2 October 2026 you can update those keyterms mid-stream.
• It runs where your compliance team needs it to. There are dedicated EU (api.eu.deepgram.com) and India (generally available 15 September 2026) endpoints, and Enterprise customers can self-host the models in their own VPC or on-premises on NVIDIA GPUs. Deepgram signs BAAs for Enterprise customers handling ePHI.
• The large contact-center vendors already buy it. Genesys runs Deepgram speech recognition inside its Agentic Virtual Agent, and Deepgram lists Five9 as a customer for self-service containment and alphanumeric accuracy. If your CCaaS already uses Deepgram, a direct contract gives you the same engine with your own keyterms and data terms.

Pricing reality

Pay-as-you-go starts with $200 of free credit, needs no card, and has no minimum. List rates as of 19 October 2026:

• Nova-3 streaming: $0.0048/min monolingual (about $0.29/hr) and $0.0058/min multilingual. Deepgram marks these streaming rates as limited-time promotional pricing, so budget for them to rise.
• Nova-3 pre-recorded: $0.0043/min monolingual.
• Flux streaming: $0.0065/min English (about $0.39/hr) and $0.0078/min multilingual.
• Add-ons: streaming diarization and redaction cost $0.0020/min each, and keyterm prompting costs $0.0013/min. A streaming call with all three runs about $0.0101/min on Nova-3, roughly double the headline rate.
• TTS: Aura-2 costs $0.030 per 1,000 characters and Flux TTS costs $0.045.
• Voice Agent API: $4.50/hr ($0.075/min) on the standard tier, $3.00/hr if you bring your own LLM and TTS, and $9.75/hr on the advanced tier. Telephony is billed separately by your carrier.

Growth plans start at $4,000 a year in prepaid credit and cut Nova-3 rates by roughly 12–17%. Overage on Growth bills at the Growth rate plus 10%. In practice, a support team transcribing 100,000 streaming minutes a month pays roughly $480–$1,010 a month at list price, depending on add-ons. Speech-to-text is rarely the expensive line in a voice-agent bill. The LLM and the voice cost more.

Best for

Support-engineering and CX platform leads who build or own a voice agent on Vapi, Retell AI, LiveKit, or Pipecat and want to choose the transcriber themselves, especially where callers read out account numbers, drug names, or product codes. It also suits RevOps teams building in-house call transcription for coaching or CRM write-back, who want the raw speech layer rather than a Gong seat.

Skip it if nobody on your team will write code against an API. Deepgram is a component, not a finished agent.

Versus the alternatives

The incumbents in speech-to-text are the hyperscalers (Google Cloud Speech-to-Text, Amazon Transcribe, Azure AI Speech) and AssemblyAI. Pick a hyperscaler when procurement requires you to stay inside an existing cloud contract and latency is not decisive. Pick AssemblyAI when price per hour drives the decision. Its Universal-Streaming model lists at $0.15/hr, about half of Nova-3’s streaming rate, though it bills on WebSocket session time rather than audio sent, so idle connections count. Its Voice Agent API is also $4.50/hr all-in, the same headline as Deepgram’s.

The fastest-growing entrant is ElevenLabs, which now sells Scribe speech-to-text next to the voices it is known for. Pick it when one vendor for both ears and voice matters more than per-minute cost, or when voice quality is what customers will judge.

Pick Deepgram when turn-taking, domain vocabulary, or self-hosting decides the project. If you want none of this in-house, buy a managed agent (PolyAI, Sierra) and let the vendor pick the speech stack.

Watch-outs

• The streaming prices are promotional. Guard: model your business case at 1.5× today’s Nova-3 and Flux streaming rates, and if volume justifies a Growth commitment, lock the rate in the order form.
• Accuracy and latency claims come from Deepgram’s own benchmarks. Guard: before you sign, run 200–500 of your own recorded calls through Nova-3, Flux, and one rival, and score word error rate on the entities that matter: account numbers, product names, and addresses.
• Flux TTS is English-only at launch. Guard: if any queue serves non-English callers, keep Aura-2 or a third-party voice on those queues, and confirm language coverage in writing before you commit to a single TTS vendor.
• Self-hosting needs an Enterprise contract and your own GPUs, and Whisper was removed from the self-hosted engine on 1 October 2026. Guard: price GPU capacity and ops headcount alongside the license, and migrate any Whisper workloads to Nova-3 before you upgrade.