Deepgram Speech to Text: Models, Benchmarks, Pricing and When to Use It

Deepgram Speech to Text: Models, Benchmarks, Pricing and When to Use It

Deepgram speech to text is a self-hosted-capable ASR API built for real-time voice applications, offering the Flux conversational model with built-in end-of-turn detection and the Nova-3 family for general transcription. It suits teams building phone or web voice agents in English and 30+ other languages who need sub-300ms streaming at commodity per-minute pricing. It is a weaker fit for deep Indic-language or code-mixed workloads.

TL;DR

• Best for: English-first real-time voice agents where turn-taking latency is the bottleneck — contact centres, outbound calling, in-app voice.
• Models: Flux English, Flux Multilingual, Nova-3 Monolingual, Nova-3 Multilingual, Nova-3 Medical. Nova-2, Enhanced and Base remain available as legacy.
• Latency: Flux posts median end-of-turn detection under 300ms (Coval, April 2026), which removes the need for a separate VAD.
• Price: Nova-3 Monolingual streaming at $0.0048/min ($0.29/hr) on Pay-As-You-Go — but that is a promotional rate against a $0.0077/min list price (Deepgram pricing page, verified July 2026 — check the vendor's pricing page before relying on this).
• Honest limitation: Numeral formatting is not supported for Hindi or Japanese on Nova-3 Multilingual (Deepgram changelog, May 2026). If your agent collects phone numbers, OTPs or rupee amounts in Hindi, this is a real problem, and it is the single biggest reason Indian teams end up swapping the STT component out.

What is Deepgram speech to text?

Deepgram speech to text is a commercial ASR platform that trains its own models rather than wrapping Whisper. Founded in 2015, it sells streaming and pre-recorded transcription over REST and WebSocket, with cloud, on-premises and private-cloud deployment.

The distinguishing architectural decision is that Deepgram built a model specifically for conversation rather than adapting a batch transcription model to streaming. Most providers do the latter. Deepgram's docs describe Flux as "the first conversational speech recognition model built specifically for voice agents," with model-integrated end-of-turn detection.

That matters more than WER for agent builders. A voice agent's perceived latency is dominated by how long it waits to be sure you stopped talking, not by how fast it transcribes.

Which Deepgram STT models are available?

Five models are current. Three are legacy but still callable.

Flux English

The conversational model for English voice agents. It handles turn-taking inside the model, so you drop the external VAD from your pipeline. Deepgram positions it at Nova-3-level accuracy with turn detection added.

Use it when the agent is English-only and you are optimising response latency.

Flux Multilingual

Released 29 April 2026. Extends Flux's integrated end-of-turn detection to multilingual audio. Priced above Flux English.

Use it for agents that need turn-taking intelligence across languages, accepting a higher per-minute rate.

Nova-3 Monolingual

Released 11 February 2025, expanded steadily since. Coverage grew through 2025–2026 to 30+ languages, with Gujarati added alongside accuracy improvements for Bengali, Marathi, Tamil and Telugu (Deepgram, May 2026).

Use it for single-language transcription — call recording, analytics, captioning — where you know the language up front.

Nova-3 Multilingual

Ten languages with real-time code-switching: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch (Deepgram, February 2026). Hindi is the only Indic language in this set.

Use it when speakers switch language mid-sentence and you cannot pin the language per call.

Nova-3 Medical

A batch model tuned for clinical terminology, upgraded with improved medical term recognition (Deepgram changelog, 2026). Batch only — not a streaming option for live clinical agents.

Legacy: Nova-2, Enhanced, Base, Deepgram Whisper Cloud

Still available. Deepgram Whisper Cloud is a managed Whisper deployment, and Deepgram's own docs state Whisper models "are less scalable than all other Deepgram models due to their inherent model architecture." Concurrency is capped at 5 on both Pay-As-You-Go and Growth, versus 150–225 for the WebSocket API. Treat it as a compatibility path, not a production one.

What features does Deepgram STT actually give you?

Accuracy and formatting. Smart Formatting handles punctuation, casing, dates and currency, and is included at no extra cost. Keyterm Prompting injects domain vocabulary — up to 100 terms on monolingual models, and up to 500 tokens (roughly 100 words) on Nova-3 Multilingual. No retraining required.

Languages and code-switching. 45+ languages across the Nova family. Real-time code-switching is limited to the ten Nova-3 Multilingual languages. Indic coverage in monolingual mode spans Hindi, Bengali, Marathi, Tamil, Telugu and Gujarati.

Real-time behaviour. WebSocket streaming with interim results. Flux's model-integrated end-of-turn detection is the differentiator — Coval measured median EOT under 300ms and estimated it saves 200–600ms of agent response time versus an STT-plus-external-VAD pipeline (Coval, April 2026). Configurable turn-taking dynamics let you tune eagerness against interruption tolerance.

Compliance and data residency. SOC 2 Type 1 and Type 2, HIPAA with BAAs for Enterprise, GDPR with a dedicated EU endpoint at api.eu.deepgram.com, CCPA and PCI. Self-hosted deployment is documented and supported. There is no India data residency option published.

Developer experience. Billing is per second of audio, not rounded to 15-second increments as some cloud providers do. On a workload of short utterances that difference is material. $200 in free credit, no card required.

Benchmarks — how accurate is Deepgram really?

Deepgram's own numbers and the independent numbers disagree by a factor of two to three. Both are worth knowing, and the gap is the story.

Two caveats we are not going to paper over. First, the two independent Artificial Analysis figures we found — 12.8% and 18.3% — conflict, and both reach us through third parties rather than from Artificial Analysis directly. We could not confirm either number at source. Second, Nova-3 does not appear on the Hugging Face Open ASR Leaderboard at all, because the leaderboard evaluates open-weight models and Nova-3 is closed. Anyone showing you a chart with Nova-3 ranked on that leaderboard has made it up.

What these numbers do not tell you. Telephony audio is 8kHz narrowband, heavily compressed, often with packet loss. Almost every public ASR benchmark is 16kHz clean read speech recorded on decent microphones. A model can win on Common Voice and fall apart on a GSM call from a two-wheeler in Bengaluru traffic. Deepgram's internal set at least includes call-centre audio, which is closer to reality than LibriSpeech, but it is still Deepgram's set, scored by Deepgram.

WER on your own call recordings is the only number that decides anything. Pull 100 real calls, transcribe them by hand, and measure. Everything above is a shortlist-generating device, not a decision.

How much does Deepgram cost?

All rates below are Pay-As-You-Go streaming, from Deepgram's pricing page, verified July 2026 — check the vendor's pricing page before relying on this.

The Growth plan ($4,000+/year prepaid) cuts these by roughly 12–15%. Enterprise is quote-based.

The headline rate is not the bill. Add-ons are priced separately:

Run Nova-3 Monolingual with diarization and redaction and you are at $0.0088/min — an 83% increase over the base line item. Budget the features you will actually ship, not the marketing rate.

One risk to model explicitly: Deepgram's pricing page states these are limited-time promotional rates on streaming. The list prices sit 30–60% higher. If your unit economics only work at $0.29/hr, you have an unhedged exposure.

Deepgram also sells a bundled Voice Agent API at $0.075/min ($4.50/hr) covering STT, LLM orchestration and TTS. That is a different product from component-level composition, and you pay for the convenience.

What is Deepgram best used for?

• Outbound collections calling for an NBFC — English and Hindi, high volume, where per-second billing on short "not interested" calls compounds into real savings.
• COD confirmation calls for a D2C brand in tier-2 India — but only in English or with Hindi numerals handled downstream. See the limitation below.
• US healthcare call-centre QA — Nova-3 Medical in batch, plus HIPAA BAAs and PII redaction, over recorded patient-access calls.
• Live agent-assist for a SaaS support desk — Flux English streaming into an LLM that surfaces answers to a human agent mid-call.
• Post-call analytics for an insurance contact centre in the EU — batch Nova-3 with diarization through the api.eu.deepgram.com endpoint for GDPR residency.
• In-app voice search for a fintech product — short utterances, where per-second billing and sub-300ms turn detection both pay off.

When should you use Deepgram?

• Your agent is English-first and response latency is the thing users complain about. Flux's integrated end-of-turn detection is the strongest answer available for that specific problem.
• You need on-premises or private-cloud deployment with a real support contract behind it.
• Your traffic is high-volume short utterances, where per-second billing beats competitors who round up.
• You are HIPAA- or PCI-bound and want a vendor that signs BAAs and holds SOC 2 Type 2.

When should you not use Deepgram?

• Your agent operates in Indic languages beyond Hindi, or in heavy Hinglish code-mixing. Nova-3 Multilingual's real-time code-switching covers Hindi only among Indic languages, and numeral formatting is unsupported for Hindi. Use Sarvam instead — Saaras v3 covers 22 Indian languages with native code-mixed handling and reports 19.31% WER on IndicVoices (Sarvam, 2026 — vendor claim, no independent Indic leaderboard exists to check it against).
• You need maximum language breadth in real time. 45+ languages is mid-pack now. Use ElevenLabs Scribe v2 Realtime (90+ languages, ~150ms) or Gladia Solaria-1 (100+ languages, ~103ms partials) if breadth is the constraint.
• You are transcribing large archives offline and cost dominates. Deepgram batch is cheap but not the cheapest. Groq runs whisper-large-v3-turbo at $0.04/hr, roughly 7× cheaper, at 216× real-time (Artificial Analysis).
• You want open weights with no licensing conversation. Deepgram is closed. Mistral Voxtral Realtime ships under Apache 2.0 with sub-200ms configurable latency and runs on a single consumer GPU.

How does Deepgram compare to the alternatives?

All figures verified July 2026 from vendor pricing pages except ElevenLabs, whose per-minute STT rate we could not confirm at source.

How do you use Deepgram on Telenow?

Deepgram is live on Telenow as a selectable STT component. Open the agent builder, go to the Speech-to-Text step, and pick Deepgram from the Deepgram STT integration page. Choose your model — Flux English for latency-sensitive English agents, Nova-3 Multilingual where code-switching matters — and toggle diarization or redaction only if you need them, since both bill separately.

Your cost line itemises the way it should. A one-hour English agent call on Flux English shows Deepgram STT at its own rate, your chosen LLM at its own rate, TTS at its own rate, and carrier minutes separately. No blended per-minute number hiding the markup. Compare against the full component pricing breakdown.

Swapping later does not mean rebuilding. If Hindi numeral formatting becomes a blocker on COD confirmation flows, change the STT component to Sarvam from the same dropdown. Prompts, tools, call logic and carrier config stay exactly as they are. That is the entire reason Telenow is composed of components rather than sold as a bundle — see the full STT provider comparison or the voice AI platform overview.

Build an agent on Deepgram and measure WER on your own calls before you commit. That is the only benchmark that matters.

FAQ

How much does Deepgram speech to text cost per hour? Nova-3 Monolingual streaming costs $0.0048 per minute, roughly $0.29 per hour, on Pay-As-You-Go as of July 2026. That is a promotional rate against a $0.0077 per minute list price. Diarization and redaction each add $0.0020 per minute.

What is the difference between Deepgram Flux and Nova-3? Nova-3 is a general transcription model. Flux is a conversational model with model-integrated end-of-turn detection, built for voice agents, at Nova-3-level accuracy. Flux removes the need for a separate voice activity detection component in your pipeline.

Does Deepgram support Hindi and other Indian languages? Yes. Nova-3 Multilingual includes Hindi with real-time code-switching. Nova-3 Monolingual additionally supports Bengali, Marathi, Tamil, Telugu and Gujarati. However, numeral formatting is not supported for Hindi on Nova-3 Multilingual as of May 2026.

Is Deepgram more accurate than Whisper? Deepgram claims a 36% lower word error rate than Whisper on selected internal datasets. That is a vendor benchmark. On Artificial Analysis, third-party reports place Nova-3 between 12.8% and 18.3% WER, above OpenAI's GPT-4o Transcribe at roughly 8.9%.

Can Deepgram be self-hosted? Yes. Deepgram documents self-hosted deployment for on-premises and private-cloud environments, and offers a dedicated EU endpoint for GDPR data residency. Self-hosting requires an Enterprise agreement. There is no published India data residency option.

What is Deepgram's latency for voice agents? Flux posts median end-of-turn detection under 300 milliseconds, measured independently by Coval in April 2026. Coval estimated this saves 200 to 600 milliseconds of agent response time compared with pairing a standard STT model with an external VAD.

Does Deepgram charge extra for speaker diarization? Yes. Speaker diarization costs $0.0020 per minute on Pay-As-You-Go and $0.0017 per minute on Growth, on top of the base model rate. Smart Formatting is included at no extra cost. Keyterm Prompting and PII redaction are also billed separately.

Is Deepgram or AssemblyAI better for voice agents? Deepgram's Flux wins on turn-taking latency through in-model end-of-turn detection. AssemblyAI's Universal-3.5 Pro Realtime covers 18 languages with context carryover. AssemblyAI bills on session duration rather than audio processed, which can produce surprising invoices if sockets are left open.

---

3. JSON-LD

[
  {
    "@context": "https://schema.org",
    "@type": "Article",
    "headline": "Deepgram Speech to Text: Models, Benchmarks, Pricing and When to Use It",
    "description": "Deepgram speech to text explained: Flux, Nova-3, Nova-3 Medical, verified pricing, independent WER benchmarks, and when to pick a different provider.",
    "url": "https://telenow.ai/blog/deepgram-speech-to-text",
    "datePublished": "2026-07-18",
    "dateModified": "2026-07-18",
    "author": { "@type": "Organization", "name": "Telenow", "url": "https://telenow.ai" },
    "publisher": {
      "@type": "Organization",
      "name": "Telenow",
      "url": "https://telenow.ai",
      "logo": { "@type": "ImageObject", "url": "https://telenow.ai/logo.png" }
    },
    "mainEntityOfPage": { "@type": "WebPage", "@id": "https://telenow.ai/blog/deepgram-speech-to-text" },
    "about": [
      { "@type": "SoftwareApplication", "name": "Deepgram" },
      { "@type": "Thing", "name": "Speech recognition" }
    ]
  },
  {
    "@context": "https://schema.org",
    "@type": "SoftwareApplication",
    "name": "Deepgram Speech-to-Text API",
    "applicationCategory": "DeveloperApplication",
    "applicationSubCategory": "Speech Recognition API",
    "operatingSystem": "Cloud, Self-hosted",
    "url": "https://deepgram.com",
    "softwareVersion": "Nova-3, Flux",
    "featureList": [
      "Real-time WebSocket streaming transcription",
      "Model-integrated end-of-turn detection (Flux)",
      "Speaker diarization",
      "PII redaction",
      "Keyterm prompting",
      "Real-time code-switching across 10 languages",
      "Self-hosted and private cloud deployment"
    ],
    "offers": {
      "@type": "Offer",
      "price": "0.0048",
      "priceCurrency": "USD",
      "description": "Nova-3 Monolingual streaming, per minute, Pay-As-You-Go promotional rate, verified July 2026",
      "url": "https://deepgram.com/pricing"
    },
    "provider": { "@type": "Organization", "name": "Deepgram, Inc." }
  },
  {
    "@context": "https://schema.org",
    "@type": "FAQPage",
    "mainEntity": [
      {
        "@type": "Question",
        "name": "How much does Deepgram speech to text cost per hour?",
        "acceptedAnswer": { "@type": "Answer", "text": "Nova-3 Monolingual streaming costs $0.0048 per minute, roughly $0.29 per hour, on Pay-As-You-Go as of July 2026. That is a promotional rate against a $0.0077 per minute list price. Diarization and redaction each add $0.0020 per minute." }
      },
      {
        "@type": "Question",
        "name": "What is the difference between Deepgram Flux and Nova-3?",
        "acceptedAnswer": { "@type": "Answer", "text": "Nova-3 is a general transcription model. Flux is a conversational model with model-integrated end-of-turn detection, built for voice agents, at Nova-3-level accuracy. Flux removes the need for a separate voice activity detection component in your pipeline." }
      },
      {
        "@type": "Question",
        "name": "Does Deepgram support Hindi and other Indian languages?",
        "acceptedAnswer": { "@type": "Answer", "text": "Yes. Nova-3 Multilingual includes Hindi with real-time code-switching. Nova-3 Monolingual additionally supports Bengali, Marathi, Tamil, Telugu and Gujarati. However, numeral formatting is not supported for Hindi on Nova-3 Multilingual as of May 2026." }
      },
      {
        "@type": "Question",
        "name": "Is Deepgram more accurate than Whisper?",
        "acceptedAnswer": { "@type": "Answer", "text": "Deepgram claims a 36% lower word error rate than Whisper on selected internal datasets. That is a vendor benchmark. On Artificial Analysis, third-party reports place Nova-3 between 12.8% and 18.3% WER, above OpenAI's GPT-4o Transcribe at roughly 8.9%." }
      },
      {
        "@type": "Question",
        "name": "Can Deepgram be self-hosted?",
        "acceptedAnswer": { "@type": "Answer", "text": "Yes. Deepgram documents self-hosted deployment for on-premises and private-cloud environments, and offers a dedicated EU endpoint for GDPR data residency. Self-hosting requires an Enterprise agreement. There is no published India data residency option." }
      },
      {
        "@type": "Question",
        "name": "What is Deepgram's latency for voice agents?",
        "acceptedAnswer": { "@type": "Answer", "text": "Flux posts median end-of-turn detection under 300 milliseconds, measured independently by Coval in April 2026. Coval estimated this saves 200 to 600 milliseconds of agent response time compared with pairing a standard STT model with an external VAD." }
      },
      {
        "@type": "Question",
        "name": "Does Deepgram charge extra for speaker diarization?",
        "acceptedAnswer": { "@type": "Answer", "text": "Yes. Speaker diarization costs $0.0020 per minute on Pay-As-You-Go and $0.0017 per minute on Growth, on top of the base model rate. Smart Formatting is included at no extra cost. Keyterm Prompting and PII redaction are also billed separately." }
      },
      {
        "@type": "Question",
        "name": "Is Deepgram or AssemblyAI better for voice agents?",
        "acceptedAnswer": { "@type": "Answer", "text": "Deepgram's Flux wins on turn-taking latency through in-model end-of-turn detection. AssemblyAI's Universal-3.5 Pro Realtime covers 18 languages with context carryover. AssemblyAI bills on session duration rather than audio processed, which can produce surprising invoices if sockets are left open." }
      }
    ]
  }
]

---

4. Internal links used

External links: deepgram.com/pricing · developers.deepgram.com/docs/models-languages-overview · developers.deepgram.com/changelog · coval.ai (independent benchmark) · artificialanalysis.ai/speech-to-text

---

5. Image concepts

1. Pipeline diagram. Two stacked rows: a conventional pipeline (STT → external VAD → LLM → TTS) above a Flux pipeline (Flux with integrated EOT → LLM → TTS), with the 200–600ms saving annotated on the gap. Alt: "Deepgram speech to text pipeline showing Flux integrated end-of-turn detection replacing a separate VAD component"

2. Stacked cost bar. One audio hour on Nova-3 Monolingual, base rate at the bottom and diarization, redaction and keyterm prompting stacked above, with the $0.29/hr and $0.53/hr totals labelled. Alt: "Deepgram speech to text pricing breakdown showing base streaming rate plus diarization, redaction and keyterm add-ons per hour"

3. Benchmark honesty chart. Side-by-side bars: Deepgram's internal 5.26% WER against the 12.8% and 18.3% third-party Artificial Analysis figures, each bar labelled vendor or independent. Alt: "Deepgram speech to text accuracy comparison between vendor-reported WER and independent Artificial Analysis benchmark results"