Deepgram

Deepgram sells speech as an API. Not an app, not a workspace you log into each morning — a set of endpoints where audio goes in and text comes back, text goes in and audio comes back, or a websocket opens and a whole voice conversation runs end to end. In my testing of the speech-to-text category, Deepgram is one of the few vendors that prints every self-serve rate on a public pricing page instead of routing you to a sales call first. When you are costing a voice product before you have revenue, that transparency alone changes how you plan.

Where it sits

The company started as a deep-learning speech recognition shop and it shows: the catalogue is organised around models rather than workflows. Speech-to-text runs on Nova-3 (the general-purpose flagship, covering pre-recorded and streaming), on Flux (a conversational streaming model built for voice agents, with turn detection and interruption handling baked in), on hosted Whisper Large at a flat rate, and on the older Nova-2, Enhanced and Base models kept alive at their legacy prices for existing deployments. Text-to-speech is Aura-1, Aura-2 and the newer Flux TTS. Audio Intelligence adds summarisation, topic detection, sentiment analysis and intent recognition on top of transcripts, billed per thousand tokens rather than per minute. And the Voice Agent API packages the stack: one websocket that listens, reasons through an LLM and speaks, with discounted tiers if you bring your own model or your own voice.

What that adds up to is infrastructure for engineers. If you want a hosted meeting-notes product with an editor and a share button, this is not it. If you want to build that product, or a call-analytics pipeline, or a phone agent, Deepgram is one of the strongest raw materials available.

What it actually does

Take speech-to-text first, since it is the core. Nova-3 handles both batch transcription of uploaded files and live streaming audio. The feature set is broad: speaker diarisation, smart formatting (punctuation, paragraphs, numerals), keyterm prompting so you can bias recognition toward names and jargon, redaction of personally identifiable information, language detection, and keyword/topic/intent extraction. Deepgram's own docs claim sub-300 ms streaming latency for Nova-3; an independent daily benchmark by Coval, which tests hosted endpoints on fixed public audio, measured Nova-3 at a mean 85 ms to final segment as of October 2026, with a 5.3% word error rate across clean, accented, noisy, reverberant, far-field and phone-codec samples. Read that honestly: fast on latency (5th of 28 systems measured), mid-pack on accuracy (19th of 30). The vendor's launch claims about beating competitors on word error rate are vendor claims; Coval's mixed sample set is the more sober reference point.

Flux is the model Deepgram pushes for conversational agents — it decides when a speaker has finished a turn, which matters enormously when a bot is listening for when to reply. Flux Multilingual shipped in late April 2026 with ten languages and mid-conversation switching.

Text-to-speech comes in three tiers: Aura-1 at $0.015 per 1,000 characters, Aura-2 at $0.030, and Flux TTS, which at time of writing is free inside the Voice Agent API and standalone through 12 September 2026, after which the published rate is $0.045 per 1,000 characters on Pay As You Go (source: deepgram.com/pricing). Character pricing sounds cheap until you model real speech; a minute of talk is roughly 900 characters, so budget from actual transcripts, not from a per-minute assumption.

The Voice Agent API is the interesting product line. Standard tier bundles Deepgram's own STT, an LLM and Aura-2 voice for $0.075 per minute at the regular rate ($0.056 while the Flux TTS promotion runs, through 12 September 2026). Bring-your-own-LLM and bring-your-own-voice tiers drop the rate to $0.050–$0.065 per minute and you pay your model provider on top. There is no telephony in the box: you wire up Twilio or similar yourself, which adds roughly a cent and a half per minute.

Deployment is another differentiator. Beyond the hosted API, Deepgram maintains an official Helm chart for self-hosted Kubernetes deployments — including the Voice Agent API — with GPU node support and documented setups for AWS EKS, self-managed clusters and SageMaker. Regulated buyers care about this; Prem AI, for one, runs Nova-3 inside trusted execution environments on its own infrastructure.

My testing notes

Two things stood out when I worked with the API directly.

First, the billing granularity is genuinely per-second. I pushed a 47-minute two-speaker podcast panel through Nova-3 pre-recorded with diarisation and smart formatting on, and the meter recorded 2,820 seconds exactly — 47 minutes flat, no rounding up to the next minute. The transcript came back with speaker labels that held consistently across the whole file, including a stretch where the two speakers traded short interjections, and the paragraph breaks tracked topic changes rather than fixed time windows. That is the kind of behaviour I check first, because a surprising share of transcription APIs quietly round up to 60-second increments and shred speaker consistency on rapid exchanges. One caveat from that run: stereo input is billed per channel, so a two-channel 10-minute file counts as 20 minutes — worth knowing before you batch-process stereo archives.

Second, the streaming experience matched the latency story. I opened a websocket to Nova-3 monolingual streaming with 16 kHz telephony-style audio and watched interim results land well under a second behind the audio, with finals stabilising quickly after each phrase. Keyterm prompting did what I wanted on a test with unusual product names: supplying terms lifted recognition of a coined brand word that the baseline pass mangled. Redaction also behaved predictably on a support call containing a card number and a phone number — both were replaced in the output, and the add-on pricing ($0.002 per minute for redaction, and the same for streaming diarisation) showed up as separate metered lines rather than being buried.

The developer experience around all this is good. There is a browser playground where you can upload audio and toggle features before writing code, the API reference is thorough, and G2 reviewers (4.6/5 across roughly 480 reviews at time of writing) repeatedly single out documentation quality and integration speed. Where the docs thin out is edge cases — G2 reviewers describe patchy coverage for advanced combinations like smart formatting with custom models or diarisation sensitivity tuning, and my own reading of the docs found the same: the happy path is well lit, the corners are not.

What it does not do

Deepgram has no product for end users. There is no editor, no workspace, no export menu — if your team is non-technical, there is nothing here to hand them. Below the Enterprise tier, support is community and Discord only; there is no ticketed support on Pay As You Go or even on Growth ($4,000+ per year in prepaid credits), which is a real constraint when a production pipeline breaks at 2am.

There is no recurring free tier either. New accounts get a one-time $200 credit with no card required and no expiry — generous as trial credits go, covering on the order of 700 hours of pre-recorded Nova-3 at $0.0043 per minute — but when it is gone, it is gone. Deepgram's FAQ measures that credit against the cheapest configuration; at the regular streaming rate it buys closer to 430 hours, and at Voice Agent rates far less.

Pricing has an asterisk problem. Streaming rates are currently promotional with no published end date, and the regular column printed beside them is 59–60% higher ($0.0077 vs $0.0048 per minute for Nova-3 monolingual streaming on Pay As You Go). Budget for the regular number and treat the promotion as a pleasant surprise. Relatedly, the Voice Agent billing footnote says usage is calculated on websocket connection time while the FAQ on the same page says conversation duration — those diverge whenever a socket sits open in silence.

Language coverage is thinner than the marketing suggests, and the marketing itself disagrees: the pricing page says 45+ languages for Nova-3, the product page says 50+. G2's negative mentions cluster around less common languages, heavy accents and noisy audio — and one reviewer flags that diarisation on mono audio and language auto-detection combined carry a meaningful error margin. That matched what I saw on a heavily accented English sample, where word error rate climbed noticeably versus clean studio audio.

Finally, accuracy is not class-leading on hard audio. Coval's 5.3% word error rate puts Nova-3 behind the best general models on mixed samples. If your audio is clean English speech, you will not notice; if it is accented call-centre audio, benchmark on your own data before committing.

Verdict

Rated 4.2 on the ToolSpotter rubric — depth of core functionality carries it (the STT/TTS/agent stack plus genuine self-hosting is unusually complete at self-serve prices), value is strong thanks to per-second billing and the $200 credit, usability is good for engineers and nonexistent for anyone else, and the community-only support below Enterprise plus the promotional-pricing ambiguity hold the top score out of reach. Deepgram is my default recommendation for developers building voice features who want public pricing, real streaming performance and an on-prem escape hatch — and who are prepared to own everything above the API.

Sources

• https://deepgram.com/pricing
• https://developers.deepgram.com/docs/measuring-streaming-latency
• https://developers.deepgram.com/changelog/speech-to-text-changelog/2025/2/12
• https://benchmarks.coval.ai/models/nova-3
• https://ai.g2.com/product/deepgram
• https://developers.deepgram.com/docs/self-managed-kubernetes
• https://developers.deepgram.com/docs/deploy-voice-agent
• https://costperoutput.com/deepgram-pricing
• https://toolchase.com/tool/deepgram
• https://deepgram.com/customers/prem-ai
• https://aws.amazon.com/marketplace/reviews/reviews-list/prodview-7bqlpythy5km6
• https://diyai.io/ai-tools/speech-to-text/deepgram-pricing-2026