Deepgram Review 2026: STT Accuracy, Per-Minute Pricing & Verdict

Vendor

Deepgram, Inc.

Category

Voice AI / STT API

Pricing Model

Pay-as-you-go + Annual

Free Credit

$200 (no card)

Founded

2015

Headquarters

San Francisco, CA

Latest Models

Nova-3, Flux, Aura-2

Best For

Real-time voice apps

Editorial Score Card

Our overall score is an editorial judgement based on the six weighted dimensions below and the scoring framework on our methodology page. It is not a user rating or an aggregate of third-party reviews. Each dimension is justified underneath the card.

How we scored each dimension

• Accuracy (9.0): Deepgram's own Nova-3 benchmarks report a 5.26% median word error rate (WER) on pre-recorded audio and 6.84% on streaming audio. These are vendor-reported numbers, not a neutral third-party test, so we score high but not perfect — real-world WER varies by accent, domain, and audio quality, and independent results may differ.
• Latency (9.4): Real-time streaming with sub-second partial transcripts is Deepgram's signature strength, and the new Flux model adds context-aware turn detection that Deepgram measures at 200-600ms faster agent responses than a stitched-together pipeline. This is the dimension where Deepgram most clearly leads.
• Pricing (8.3): Per-minute rates are transparent, published, and low relative to the big cloud providers, and the $200 free credit is genuinely usable. Points come off because add-ons (diarization, redaction, keyterm prompting) are billed separately and the deepest discounts require an annual Growth commitment.
• Developer experience (9.0): First-class SDKs, WebSocket streaming, a live API playground, and thorough documentation make this one of the smoother onboarding experiences in voice AI.
• Support (7.8): Self-serve tiers rely on docs and community channels; dedicated support, SLAs and on-prem/self-hosted options are reserved for Enterprise contracts. There is no on-demand premium support tier for pay-as-you-go users.
• Overall (8.8): A weighted blend that reflects a best-in-class real-time engine and developer experience, tempered by add-on costs and a maturing (but still younger) voice-agent and audio-intelligence layer relative to specialists.

Deepgram Pricing 2026 (Verified)

Deepgram uses a consumption-based model: you pay per minute of audio processed (for speech-to-text) or per 1,000 characters (for text-to-speech). The figures below were taken from Deepgram's public pricing page in July 2026 and reflect the pay-as-you-go (PAYG) rates, with Growth (prepaid annual) rates noted where they differ. Always confirm current pricing directly with Deepgram before committing spend.

The core plan structure has three tiers: Pay As You Go (start free with a one-time $200 credit, no credit card, no minimums and no expiration on the credit); Growth (prepaid annual credits, published as starting from roughly $4,000+/year, with savings of up to about 20% versus PAYG); and Enterprise (custom volume pricing, SLAs, and deployment options negotiated with sales).

Voice Agent API: Deepgram's hosted voice-agent service is billed per WebSocket minute. Published PAYG rates are $0.075/min for the Standard tier and $0.163/min for the Advanced tier. Bring-your-own options lower the cost: BYO TTS is about $0.065/min, BYO LLM about $0.059/min, and BYO LLM + TTS about $0.050/min. Growth pricing shaves a further ~10% off these (Standard drops to roughly $0.068/min). The LLM step routes to third-party providers such as OpenAI, Anthropic and Meta rather than a Deepgram-built model.

What We Like & What We Don't

Deepgram Feature Review: The Full Analysis

Nova-3: The Flagship Speech-to-Text Model

Nova-3 is Deepgram's current flagship automatic speech recognition (ASR) model and the engine most buyers will evaluate first. Deepgram positions it as a step change over Nova-2, and the company's published benchmarks are aggressive: a 5.26% median word error rate on pre-recorded audio (which Deepgram frames as a 47.4% improvement over the next-best competitor it tested, at roughly 10% WER) and a 6.84% median WER on streaming audio (a claimed 54.2% improvement over a next-best competitor at 14.92%). We report those numbers because they are the ones Deepgram publishes, but buyers should treat them as vendor-reported marketing benchmarks rather than neutral third-party results. Word error rate is highly sensitive to the test set — accent, background noise, domain jargon, microphone quality and audio codec all move the number — so the only WER that matters for your decision is the one you measure on your own representative audio.

The practical strengths of Nova-3 are consistent, though. On clean and moderately noisy English audio it produces clean, well-formatted transcripts with smart formatting (automatic punctuation, capitalisation and number normalisation), word-level timestamps, and profanity filtering. It handles telephony-grade 8kHz audio — the bread and butter of contact centres — without the accuracy collapse that plagues models trained mostly on studio audio. And it exposes keyterm prompting, which lets you supply up to 100 domain-specific terms (drug names, SKUs, company jargon) to bias recognition without retraining a custom model. Deepgram cites a veterinary-domain example with a large jump in keyterm recognition; in our reading this is the single most valuable accuracy lever for specialised deployments, and it is far cheaper and faster than full fine-tuning.

Where Nova-3 genuinely separates itself is latency on streaming audio. Deepgram returns partial transcripts over a persistent WebSocket connection with sub-second responsiveness, and the model keeps inference times among the fastest in the category even with features like diarization enabled. For a live voice agent, this is the difference between a natural back-and-forth and an awkward pause-and-wait exchange. Cloud-hosted implementations of open models such as Whisper typically add meaningful latency in comparison, which is why teams building conversational products so often reach for a purpose-built streaming API.

Nova-3 Multilingual and Language Coverage

Deepgram supports transcription across dozens of languages spread over its model family, but the specific 2026 headline for Nova-3 is real-time multilingual code-switching across 10 languages: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch. Code-switching means the model can follow a speaker who switches languages mid-sentence — common in bilingual households, international support lines, and many global markets — without you having to pre-declare a single language. That is a harder problem than single-language transcription and a meaningful differentiator for global contact centres. Multilingual audio is billed at the higher multilingual streaming rate ($0.0092/min PAYG versus $0.0077/min monolingual), so budget accordingly if your traffic is genuinely mixed-language. Some lower-resource languages remain available only on older Deepgram models with reduced accuracy, so verify coverage for your exact language list before committing.

Flux: Conversational Speech Recognition for Voice Agents

The most important product development for voice-agent builders is Flux, which Deepgram launched in October 2025 and describes as the first production conversational speech recognition (CSR) model, with a Flux Multilingual variant following in 2026. Where a traditional ASR model treats speech like dictation — streaming a running transcript that some other system has to interpret — Flux fuses transcription and turn detection into a single model that understands dialogue flow. In plain terms, Flux tries to know when a speaker has actually finished a thought.

This matters because the hardest problem in voice agents is not transcription accuracy; it is turn-taking. Silence-based voice-activity-detection (VAD) systems interrupt callers who simply pause to think, or wait too long after a caller has clearly finished. Flux uses semantic context to distinguish an incomplete thought ("because…", "uh, sorry…") from a complete one ("Thanks so much."), emitting a turn-complete transcript at the right moment. Deepgram reports roughly 30% fewer false interruptions and a 200-600ms reduction in agent response latency compared with a stitched-together pipeline, with over half of turn detections landing within 500ms and P90 latency around one second. For anyone who has built a voice bot and fought endless barge-in and end-of-turn tuning, that is a compelling pitch. Flux matches Nova-3's transcription accuracy while adding the conversational layer, and at $0.0065/min for English streaming it is priced competitively with Nova-3's monolingual streaming rate.

Pre-Recorded vs. Real-Time: Two Distinct Pricing Tracks

Deepgram offers separate endpoints and pricing for pre-recorded (batch) and streaming (real-time) audio, and the gap is meaningful. On Nova-3, monolingual pre-recorded transcription is priced at $0.0048/min PAYG, while monolingual real-time streaming is $0.0077/min — the premium reflects the infrastructure cost of maintaining persistent low-latency connections. The practical implication for buyers is straightforward. If your workload is post-call analytics, meeting transcription from recordings, or media/content transcription at scale, the pre-recorded endpoint is highly cost-competitive and you should benchmark it against AWS Transcribe, Azure AI Speech, and Google. If you are building live voice agents, real-time captions, or interactive telephony, the streaming (or Flux) endpoint is where Deepgram's latency advantage justifies the higher per-minute rate.

One nuance that the base rate hides: add-ons are billed on top. Speaker diarization adds about $0.0020/min, PII redaction another ~$0.0020/min, and keyterm prompting ~$0.0013/min. A streaming deployment that needs diarization plus redaction is realistically closer to $0.012/min than the $0.0077 headline. That is still inexpensive relative to legacy vendors, but it changes the total-cost-of-ownership math at high volume, and it is a correction worth flagging because diarization is sometimes assumed to be free.

Voice Agent API: The Complete Voice AI Stack

Deepgram's Voice Agent API chains speech-to-text, an LLM step, and text-to-speech into a single managed WebSocket service. Rather than operating three separate vendor relationships and building your own orchestration and latency-optimisation layer, you can hand the whole pipeline to Deepgram and receive a single bill. Pricing is per WebSocket minute: the Standard tier is $0.075/min PAYG and the Advanced tier is $0.163/min PAYG, with bring-your-own tiers that cut cost if you already have LLM or TTS relationships — BYO TTS at roughly $0.065/min, BYO LLM at roughly $0.059/min, and BYO LLM + TTS at roughly $0.050/min. Growth pricing trims these further. Importantly, the LLM is not a Deepgram model: the service routes to third-party providers such as OpenAI, Anthropic and Meta, so your model choice (and its cost/quality trade-off) is part of the configuration.

Beyond the pipeline economics, the Voice Agent API provides the agent-level plumbing that voice applications actually need: interrupt/barge-in handling so a caller can talk over the agent, silence and end-of-turn handling (now strengthened by Flux), and support for telephony and WebRTC transport so it slots into real call infrastructure. For teams that want to ship a voice agent without assembling and tuning a multi-vendor stack, this is a genuine accelerant. For teams that already run a sophisticated orchestration layer, or that want the deepest library of pre-built agent templates and flows, dedicated platforms such as Vapi or Retell AI may still fit better — the Voice Agent API is powerful but younger and lighter on templates.

Text-to-Speech: Aura-2 Quality and Pricing

Aura-2 is Deepgram's enterprise-grade text-to-speech model, positioned for real-time voice agents where speed matters as much as timbre. It generates natural, appropriately paced speech across a library of pre-built voices selectable by name via the API, and its defining advantage is latency: within the Voice Agent API, the time from LLM output to first audio byte is very low, which is exactly what conversational applications need. On pure voice quality for long-form, expressive content, our assessment is that Aura-2 trails a specialist like ElevenLabs — the gap is most audible on complex sentence structures and sustained narration. For short conversational turns in a phone agent, the difference is far less important than the latency win.

On price, Aura-2 is $0.030 per 1,000 characters on PAYG (about $0.027 on Growth), which is double the $0.015/1k characters of the older Aura-1 model. If your use case is high-volume, quality-sensitive narration rather than real-time agents, it is worth pricing Aura-2 against ElevenLabs and other TTS specialists — and considering the BYO TTS path in the Voice Agent API if you prefer a different voice engine while keeping Deepgram's STT and orchestration.

Speaker Diarization and Audio Intelligence

Speaker diarization — labelling which words belong to which speaker — is available on Nova models, but, contrary to a common misconception, it is a paid add-on, not a free feature: expect roughly $0.0020/min on top of the base transcription rate. Diarization is valuable for meeting transcription, legal depositions, and contact-centre analytics where speaker attribution feeds downstream QA and compliance. If speaker labels are central to your product, factor the add-on into your per-minute cost from the start.

Deepgram also offers audio-intelligence features beyond raw transcription — sentiment, topic detection, intent, and summarisation. These are useful, but in our assessment they are not as deep or as accurate out of the box as AssemblyAI's audio-intelligence suite, which offers richer entity detection and more granular topic taxonomies. Teams whose primary need is post-call analytics depth (rather than real-time performance) should evaluate AssemblyAI as the analytics layer, potentially pairing it with Deepgram for the real-time components. Teams whose primary need is live, low-latency voice should stay with Deepgram and treat analytics as a secondary consideration.

Developer Experience and SDK Ecosystem

Developer experience is one of Deepgram's strongest suits. The platform ships official SDKs for Python, JavaScript/TypeScript, Go and .NET, with community libraries for other languages, and the documentation is comprehensive — interactive examples, WebSocket connection guides, and model comparison tables. A live API playground lets you test transcription in the browser before writing any code, which shortens proof-of-concept cycles considerably. In practice, a working WebSocket streaming integration is on the order of 50-100 lines of code using the official SDK.

Deepgram supports three main connection modes: real-time streaming over WebSocket, batch processing over REST, and the hosted Voice Agent API (WebSocket with additional signalling for the agent pipeline). Authentication uses API keys with optional key-level permission scoping, and enterprise security requirements such as IP allowlisting are supported. Combined with the transparent pricing and the free $200 credit, this makes Deepgram unusually easy to trial rigorously — which is exactly what we recommend before committing, precisely because accuracy claims should be validated on your own audio rather than taken from any vendor's benchmark table.

Integrations

Deepgram integrates natively or through community libraries with the telephony platforms, cloud storage, LLM providers, and voice-agent orchestration tools that voice applications typically depend on:

Best Use Cases

Who Deepgram Is Best For

Deepgram is the right choice for developer teams building voice-first applications — contact-centre AI, voice agents, real-time transcription platforms, and meeting-intelligence tools. It is API-first by design, which gives engineering teams maximum flexibility to embed speech recognition into custom products, and its latency profile is the best reason to choose it over a general cloud speech service.

It is also a strong fit for SaaS companies building speech-enabled products that need a reliable, scalable transcription backend with transparent usage-based billing. ISVs benefit from the generous free credit for development, low per-minute rates that protect margins at scale, and Enterprise SLAs for production. If real-time latency and per-minute cost predictability are your top two criteria, Deepgram should be at the top of your shortlist.

Who Should Look Elsewhere

Deepgram is not the right fit for non-technical teams wanting a ready-made transcription tool. There is no consumer app — everything is API-based. Teams who just want click-and-use meeting transcription should look at consumer tools like Otter.ai or Fireflies.ai, which are built on top of speech APIs.

Teams whose primary need is deep post-call audio analytics — rich entity extraction, custom topic taxonomies, templated call scoring — should evaluate AssemblyAI, which ships a more comprehensive out-of-the-box analytics layer. And teams that want the most opinionated voice-agent orchestration with the largest library of pre-built flows and templates may prefer a dedicated platform such as Vapi or Retell AI, potentially still running Deepgram underneath for STT.

Deepgram Alternatives

Verdict

Deepgram remains a front-runner for real-time speech-to-text and has built out a credible end-to-end voice stack with Nova-3, the conversational Flux model, Aura-2 TTS, and the hosted Voice Agent API. For any team where voice latency is a first-class requirement — contact-centre AI, voice bots, live captions, real-time transcription — Deepgram should be your first evaluation, and the free $200 credit makes that evaluation easy to run properly.

The platform scores highest on latency and developer experience, and its per-minute pricing is transparent and low. The honest caveats: accuracy claims are vendor-reported and must be validated on your own audio; add-ons like diarization and redaction meaningfully raise the effective per-minute cost; Aura-2 TTS trails specialist voices on long-form quality; and the audio-intelligence and voice-agent layers, while strong, are still catching specialists like AssemblyAI (analytics) and Vapi/Retell AI (agent orchestration). Weighed together, Deepgram earns a strong editorial score for its core mission — production-grade, low-latency speech recognition — and is our recommended starting point for developer teams building voice-first applications in 2026.

Editorial score: 8.8/10 — Highly recommended for developer teams building real-time, voice-first production applications.

Frequently Asked Questions