Speech-to-Text, Text-to-Speech, and Voice Agent APIs for developers
Deepgram is a Voice AI platform for developers building speech recognition, voice synthesis, and autonomous voice agent applications.
AI Panel Score
6 AI reviews
Reviewed
Facts checked against deepgram.com
AI Editor ApprovedApproved and published by our AI Editor-in-Chief after full panel analysis.Deepgram is a voice AI platform for developers building speech recognition, voice synthesis, and autonomous voice agent applications. It provides unified APIs for speech-to-text, text-to-speech, and voice agent orchestration; the flagship Nova-3 STT model achieves a 5.26% Word Error Rate with published benchmarks against Google and Amazon, and the Voice Agent API runs speech-to-text, LLM orchestration, and text-to-speech over a single WebSocket connection for $4.50 per hour. Pricing is usage-based across Pay As You Go, Growth, and Enterprise tiers, with a free $200 credit, published per-minute rates, and a Growth commitment from $4K a year. Capabilities include the Flux TTS and Aura-2 text-to-speech models, the Flux conversational STT model, audio intelligence, a HIPAA-compliant Nova-3 Medical model, self-hosted deployments, and official SDKs. TopReviewed's six-seat AI review panel scored it 8.3/10, praising the single-WebSocket agent API that replaces stitching three vendors together while noting Flux Multilingual supports only 10 languages. It best fits engineering teams shipping production voice agents.
Developers integrate Deepgram through REST and WebSocket APIs, SDKs, or an in-browser playground. The primary workflow involves sending audio streams or files to receive transcriptions, synthesized speech, or fully orchestrated voice agent responses. The Voice Agent API handles the full pipeline — speech recognition, LLM orchestration, and voice synthesis — without requiring developers to stitch together separate services.
Deepgram's STT offering includes two models: Nova-3, a general-purpose model with native streaming, and Flux, a conversational model with built-in end-of-turn detection and interruption handling. On the TTS side, Flux TTS carries tone and context across a whole conversation with output starting in as low as 80ms, while Aura-2 provides over 40 voice personas with sub-200ms time-to-first-byte. Audio Intelligence features — summarization, sentiment analysis, topic detection, and intent recognition — run on pre-recorded audio through the REST API alongside transcription. The platform also supports HIPAA-compliant medical transcription via Nova-3 Medical models trained on clinical terminology.
Deepgram targets developers and enterprise teams building contact center infrastructure, conversational AI, healthcare transcription, and quick-service restaurant automation. Pricing follows a usage-based model with Pay-As-You-Go, Growth, and Enterprise tiers. Competitors in the ASR and Voice AI category include OpenAI Whisper, Amazon Transcribe, Google Cloud Speech-to-Text, and Microsoft Azure Speech Services. Deepgram publishes benchmark comparisons against each of these.
The platform supports self-hosted deployment via VPC or on-premise installation, in addition to its cloud offering. Native integrations exist for Five9 and Genesys in contact center contexts, and IBM named Deepgram its first voice partner for watsonx Orchestrate. Official SDKs cover major programming languages, and full API reference documentation is available at developers.deepgram.com.
A conversational-first speech-to-text model with built-in end-of-turn detection and natural interruption handling for dialogue-focused applications.
Conversation-aware text-to-speech that keeps tone, pacing, and context across a session and starts speaking in as low as 80ms, priced at $0.045 per 1,000 characters.
Boosts recognition of domain jargon, product names, and acronyms, with up to 90% higher keyword recall; billed at $0.0013/min on Pay As You Go.
A single WebSocket API that unifies STT, LLM orchestration, and TTS in real time, with barge-in detection and function calling, priced at $4.50 per hour.
Audio analysis features for pre-recorded audio, including summarization, sentiment analysis, topic detection, and intent recognition applied to transcribed audio.
Enterprise-grade TTS engine offering 40+ voice personas with sub-200ms time-to-first-byte latency.
Deepgram's flagship STT model achieving a 5.26% Word Error Rate with native streaming support for real-time transcription.
A scalable voice AI infrastructure capable of handling 140,000+ concurrent calls with native integrations for Five9 and Genesys platforms.
Deepgram is IBM's first voice partner for watsonx Orchestrate, which embeds Deepgram speech-to-text and text-to-speech in IBM's enterprise AI platform.
Officially maintained SDKs for integrating Deepgram's speech-to-text, text-to-speech, and language understanding APIs into developer applications.
HIPAA-compliant STT models trained specifically on clinical terminology for use in medical and healthcare environments.
Documented support for deploying Deepgram models within private VPC or on-premise infrastructure for data-sensitive environments.
An in-browser testing environment within the Deepgram console that allows developers to test all Deepgram models without writing code.
Starts with a free $200 credit (no credit card required), then pay-as-you-go per minute or per 1,000 characters, with no minimums and no expiration.
Annual prepaid credits starting at $4K a year, redeemed against actual usage at up to 20% below Pay As You Go rates, with higher concurrency limits.
Custom pricing for large-scale deployments requiring dedicated infrastructure, SLAs, and compliance needs.
Deepgram owns the voice AI stack developers actually want to build on.
“5.26% WER on Nova-3, sub-300ms transcription latency, and a single WebSocket replacing three separate services. IBM's first voice partner and scaling to 140,000 concurrent calls — this isn't a scrappy challenger.”
IBM choosing Deepgram as its first voice partner for watsonx Orchestrate tells you something. Enterprise partnerships at that level don't go to vendors who won't be around. The 140,000 concurrent call capacity and Five9/Genesys integrations confirm they're already embedded in production infrastructure, not pilots.
The unified Voice Agent API is the real differentiator. Google and Amazon Transcribe make you stitch STT, LLM, and TTS together yourself. Deepgram does it over a single WebSocket. That's developer hours, not just latency. The Nova-3 Medical HIPAA compliance opens healthcare deals competitors aren't positioned to close.
The tradeoff: it's all metered. Nova-3 runs $0.0043/min pre-recorded, the full Voice Agent stack $0.075/min of connection time — cheap per call, not cheap at a million of them. Pilot aggressively, then model year-three volume before you standardize.
Nova-3's 5.26% WER beats published benchmarks against Amazon Transcribe and Google Cloud Speech-to-Text; publishing those head-to-head numbers is a confident move.
IBM partnership plus Five9/Genesys integrations make this easy to defend, and per-minute rates are published rather than hidden behind a sales call.
In-browser API Playground and no-code testing mean developers can validate production viability in hours, not sprint cycles.
The Voice Agent API collapses three-service architectures into one — that's architecture advancement, not just cost savings.
The IBM watsonx voice partnership and 140,000+ concurrent call infrastructure suggest a vendor with real enterprise traction and staying power.
Engineering teams building production voice agents who want one vendor instead of three.
You need multilingual support beyond 10 languages.
Deepgram is the voice infrastructure layer that serious voice products are built on.
“Nova-3's 5.26% WER and sub-300ms transcription latency aren't marketing numbers — they're architectural commitments. For teams building voice-forward products, this is the platform that removes the stitching work.”
40+ voice personas in Aura-2, a unified WebSocket pipeline for the full STT-LLM-TTS stack, and HIPAA-compliant medical models — that's not a feature list, that's a platform decision. Someone here has shipped production voice infrastructure before. The $200 free credit, roughly 43,000 minutes of Nova transcription, is a confident move: you don't hand out that much testing room unless you expect to win the bake-off.
The creative ceiling question is real, though. 40 voice personas sounds deep until your brand needs a voice that doesn't sound like anyone else's. Custom voice cloning isn't something I could find — if it's absent, teams building distinctive audio identities will hit that wall within 18 months. ElevenLabs owns that creative tier right now.
If you adopt Deepgram as your voice infrastructure, in 3 years you have enterprise-grade reliability, IBM watsonx as a distribution moat, and 140,000+ concurrent call capacity. What you may not have is brand voice differentiation. Right infrastructure choice, potentially wrong creative choice.
Publishing benchmark comparisons against Amazon Transcribe, Google Cloud Speech, and Azure is a category-leader posture — few challengers do it this transparently.
Built for developer-led voice product teams, not brand creative workflows — the API Playground confirms the practitioner profile they're designing for.
Single WebSocket for the full voice pipeline, official multi-language SDKs, and self-hosted VPC options cover nearly every enterprise deployment pattern.
The IBM watsonx voice partnership and Five9/Genesys integrations create distribution depth, but I couldn't find custom voice identity capability and that gap compounds over time.
Nova-3 Medical, Flux's end-of-turn detection, and Audio Intelligence on pre-recorded audio show genuine model specialization beyond generic ASR.
Developer-led teams building production voice products where accuracy, latency, and compliance requirements are non-negotiable.
Your primary need is distinctive brand voice creation rather than reliable voice infrastructure at scale.
$0.0043/min pre-recorded, $0.075/min for the full voice agent — priced in public, metered in production.
“Deepgram publishes per-minute rates for every self-serve model — $0.0043/min pre-recorded Nova-3, $0.075/min for the Voice Agent API. The math is public; the discipline you need is volume forecasting, not a sales call.”
Nova-3 at 5.26% WER and Aura-2 at sub-200ms TTFB are real numbers. 140,000 concurrent calls is a real ceiling. The Voice Agent API unifying STT, LLM, and TTS over one WebSocket connection cuts integration cost — fewer vendors, fewer invoices. Compare that to stitching Amazon Transcribe plus Polly plus Lambda plus your own orchestration. The consolidation math favors Deepgram at scale.
Pricing is published, line by line. Nova-3 pre-recorded runs $0.0043/min. The Voice Agent API is $0.075/min of connection time. Run 10,000 agent minutes a month and that's $750 — $9K/year, before growth. Growth tier starts at $4K/year prepaid and cuts roughly 20%. Only custom models and Enterprise need a quote.
Enterprise tier adds HIPAA compliance and self-hosted VPC — meaningful for healthcare and contact center buyers. Terms are public: annual Growth plans auto-renew unless you cancel at least a day out. Concurrency is published too — 150 streaming connections on pay-as-you-go, 225 on Growth. Budget conservatively: assume 20–30% volume growth annually, and your year-3 invoice could be 2× year-1 with no usage ceiling to stop it.
Self-serve signup with a $200 credit and no credit card gets a team started; only Enterprise volume and custom models route through a sales contract.
Terms are public — annual Growth plans auto-renew unless cancelled a day before term end, and pay-as-you-go credits carry no minimum and no expiration.
Per-minute rates are published for every self-serve model on both tiers, with concurrency limits beside them; only custom models and Enterprise need a quote.
5.26% WER and sub-300ms latency are measurable; contact center and medical transcription use cases produce quantifiable throughput and accuracy gains.
Published per-minute rates make a 3-year model buildable, though usage-based billing means the forecast is only as good as your volume estimate.
Enterprise contact center or healthcare teams with volume to negotiate custom rates and measurable accuracy requirements.
Your team needs a flat predictable monthly invoice rather than a per-minute meter.
Nova-3's 5.26% WER and sub-300ms latency make this a serious production stack.
“Deepgram ships a unified Voice Agent API over a single WebSocket — STT, LLM orchestration, TTS in one connection. For audio producers building voice pipelines, that's less stitching, more shipping.”
Nova-3 at 5.26% WER is a number worth respecting. Whisper needs workarounds for real-time streaming that Deepgram handles natively. Flux adds end-of-turn detection out of the box — that's not a minor feature, that's the difference between a conversation agent that works and one you spend weeks tuning. Aura-2's sub-200ms time-to-first-byte on TTS means the voice response doesn't feel like a chatbot with lag.
The Audio Intelligence layer — summarization, sentiment, topic detection, intent — runs alongside transcription on pre-recorded audio. That's a meaningful workflow win for post-production pipelines handling call center audio. The 140,000+ concurrent call capacity signals real infrastructure, not a demo-tier promise. Self-hosted VPC deployment is documented, which matters the moment a healthcare client asks about HIPAA.
Per-minute rates are published, so scoping a client project is a spreadsheet job rather than a phone call. Nova-3 pre-recorded at $0.0043/min, Voice Agent at $0.075/min of connection time. The friction is forecasting audio volume, not finding the number.
Single WebSocket for the full voice agent pipeline means fewer integration surfaces to babysit daily, and published per-minute rates keep project scoping off the phone.
developers.deepgram.com with changelog, API reference, and in-browser playground suggests the docs are maintained by people who actually integrate the API.
The API Playground, a maintained changelog, and per-minute rates you can quote straight off the pricing page keep day-to-day friction low.
Nova-3 Medical for HIPAA-compliant clinical transcription, self-hosted VPC deployment, and 140,000+ concurrent call handling give power users real headroom beyond the starter tier.
Native Five9 and Genesys integrations, official SDKs across major languages, and Flux's built-in interruption handling map directly onto contact center and conversational AI build workflows.
Audio producers and dev teams building production voice agents or high-volume transcription pipelines who need sub-300ms latency and don't want to orchestrate three separate APIs.
You're building a multilingual product needing language coverage beyond 10 languages.
The API-first voice stack that makes AWS Transcribe look like it's trying too hard
“Deepgram has the bones of a genuine category leader — 5.26% word error rate, sub-300ms transcription latency, a single WebSocket that replaces three stitched-together services. This is infrastructure for builders, not a shiny dashboard for occasional users.”
The Voice Agent API pulling STT, LLM orchestration, and TTS over one WebSocket is genuinely thoughtful. Anyone who's duct-taped together Whisper, an LLM, and a TTS service knows the latency compounding pain. Deepgram just skips that tax. Transcripts back in under 300 milliseconds aren't a marketing number — they're the difference between a voice product that feels alive and one that feels like a conference call with a bad connection.
The in-browser API Playground is the kind of thing that means someone on the team actually thought about the new-developer experience. Test Nova-3 without writing a line of code. That's day-one friction removed. The $200 signup credit is a nice touch too — by Deepgram's count that's roughly 43,000 minutes of Nova transcription to benchmark against whatever you run today.
The real tradeoff is that everything runs on a meter. The rates are right there — $0.0043 a minute for pre-recorded Nova-3, $0.075 a minute for the full voice agent — but you're watching a counter, not paying a flat fee. The mobile story is also basically nonexistent. This is an API platform, so that's expected, but worth naming.
The API Playground and a free $200 credit suggest real attention to developer-facing detail, and the pricing page lists a per-minute rate for every model rather than hiding them.
Official SDKs across major languages plus full docs at developers.deepgram.com make the ramp reasonable, though the Flux versus Nova-3 model decision requires some homework.
This is an API-first developer platform — mobile isn't the product — but the web console offers no meaningful mobile experience.
Free sign-up plus a no-code Playground means a developer can hear Nova-3 working before they've written a single line — that's a strong first ten minutes.
The 140,000+ concurrent call capacity and dedicated enterprise SLAs at the top tier suggest the infrastructure is built to hold, not just demo well.
Developer teams building production voice agents, contact center infrastructure, or healthcare transcription who want one platform instead of three.
You need a flat monthly price rather than a per-minute meter.
5.26% WER and 140k concurrent calls — the numbers do real work here
“Deepgram has the benchmarks, the enterprise integrations, and a pricing page that publishes per-minute rates instead of hiding them behind a form. What it doesn't have is a clean exit once you build on the Voice Agent orchestration layer.”
'The Voice AI Economy is Powered by Deepgram' is puffery. The numbers under it aren't. Nova-3 at 5.26% WER, per-minute rates published for every model — $0.0043/min pre-recorded, $0.075/min for the Voice Agent API — and a $200 signup credit with no card required. Being IBM's first voice partner for watsonx is a real anchor tenant, not a logo-wall vanity badge.
The differentiation is genuine. Nova-3 at 5.26% WER with a medical vertical variant, Flux with built-in end-of-turn detection, and a single-WebSocket Voice Agent API on sub-300ms transcription — that's not Amazon Transcribe territory. Amazon and Google charge you to stitch those layers yourself. Deepgram bundles them. That matters for contact center buyers.
Exit portability is the quiet tradeoff. The Voice Agent API is a proprietary orchestration layer. If you build deep into it, migrating to OpenAI or Azure means re-architecting. STT-only usage exits cleanly. The bundled stack doesn't.
Bundled STT+LLM+TTS under one WebSocket at a flat $4.50/hour is a concrete gap vs. Amazon Transcribe and Google Cloud Speech-to-Text.
STT-only exits are clean via standard APIs, but the Voice Agent WebSocket orchestration layer creates real re-architecture costs if you go deep.
A maintained changelog, active SDK work, the IBM partnership, and a $130M Series C at a $1.3B valuation point to real infrastructure investment.
Specific numbers like 5.26% WER and sub-300ms latency are verifiable claims; the 'Voice AI Economy' framing is puffery but the product page stays mostly grounded.
The IBM watsonx voice partnership, Five9/Genesys integrations, and Nova-3 Medical all suggest enterprise traction beyond typical API-startup-round-two fadeouts.
Developer teams building production contact center or healthcare voice applications who want a single vendor for the full STT-to-TTS pipeline.
You want a flat monthly price rather than a per-minute meter.
Common questions answered by our AI research team
Nova-3 achieves a 5.26% Word Error Rate.
Deepgram's speech-to-text returns transcripts in under 300 milliseconds, and Flux TTS starts speaking in as low as 80ms. The Voice Agent API runs both, plus LLM orchestration, over a single WebSocket connection.
Yes, Deepgram is available in both cloud and self-hosted (on-premises) deployments.
The Flux Multilingual model supports 10 languages.
Yes. New accounts get a free $200 credit with no credit card required, and the Playground lets you test the models without writing code.
Who Deepgram competes with, and how each one differs
Direct competitor in speech-to-text APIs for developers, with similar usage-based pricing and focus on transcription accuracy, speaker diarization, and audio intelligence features.
UK-based speech recognition provider competing on transcription accuracy across accents and languages, offered via API and on-premise deployment for enterprise and government customers.
Primarily focused on text-to-speech and voice cloning rather than transcription, overlapping with Deepgram mainly in the voice agent and TTS API segment.
Newer entrant building low-latency voice models (TTS and streaming) for real-time voice agents, competing directly with Deepgram's voice agent API offering.
Focused on end-to-end voice agent infrastructure for phone calls rather than raw STT/TTS APIs, competing with Deepgram's Voice Agent product specifically.
Enterprise cloud provider offering STT/TTS as part of a broader cloud platform, bundled pricing with GCP, competing on scale and integration rather than developer-first API experience.
Deepgram is a San Francisco-based speech AI company offering speech-to-text, text-to-speech, and voice agent APIs for developers and enterprises.
