Deepgram vs AssemblyAI in 2026: Latency, VAD & ASR TCO
Deepgram vs AssemblyAI and OpenAI Realtime ASR in 2026, compared on engineering and 36-month TCO: Flux native End-of-Turn vs Universal-3.6 Pro Realtime, Nova-3 keyterm prompting, PII audio redaction, silence billing on streaming, Silero VAD, and L4 GPU math.

On this page

Teams that scale voice agents, clinical ambient scribes, or contact-center QA past beta usually end up comparing Deepgram vs AssemblyAI (or AssemblyAI vs Deepgram, depending on which one they tried first). The questions underneath are Deepgram Nova-3 vs AssemblyAI Universal-2, speech to text API pricing, and voice agent STT latency. At 5,000 minutes a month the choice barely matters. At 1,500,000 minutes a month (25,000 hours), turn detection, jargon recall, PII redaction add-ons, and billing for silence move the 36-month bill by $201,600 to $388,800 in our worked model.
Keep the neighbouring layers separate. Vapi vs Retell covers the voice orchestration wrappers that sit above your STT, LLM, and TTS keys. LiveKit vs Agora covers WebRTC SFU media transport and TURN relays, and Modal vs Baseten covers general serverless GPU autoscaling. This article is only about Automatic Speech Recognition (ASR) engines: real-time turn detection, noisy multi-speaker batch transcription, compliance redaction, and what an owned L4 GPU cluster costs.
At Idealogic, our AI agent development, AI software development, generative AI development, and custom software development teams engineer sub-300ms speech pipelines across healthtech, fintech, and logistics and supply chain. Here we compare Deepgram (Nova-3 and Flux), AssemblyAI (Universal-2, Universal-3.5 Pro, and Universal-3.6 Pro Realtime), OpenAI (GPT-4o-transcribe and Realtime), and an owned L4 GPU cluster, then show how a Silero VAD silence gate cuts streaming cost and walk through the 36-month TCO math. Vendor prices are November 2026 list prices. Deepgram tags some as promo pricing and the plan tier (pay-as-you-go or Growth) changes every number, so recheck before you budget.
The short version
- Deepgram is the pick for voice agent turn-taking and runtime jargon. On Deepgram pricing, Nova-3 pre-recorded is $0.0048/min pay-as-you-go ($0.0042 Growth, multilingual $0.0050 to $0.0058) and streaming is $0.0077/min pay-as-you-go ($0.0065 Growth). Keyterm Prompting needs no retraining. Deepgram Flux ($0.0065/min pay-as-you-go, $0.0057 Growth) detects End-of-Turn (EoT) inside the model. HIPAA BAAs and self-hosted containers come with Enterprise agreements.
- AssemblyAI is the pick for noisy multi-speaker recordings and Audio Intelligence. On AssemblyAI pricing, Universal-2 ($0.15/hr, $0.0025/min async) and Universal-3.5 Pro ($0.21/hr, $0.0035/min async) are the batch models, and 3.5 Pro is the one AssemblyAI positions for crosstalk-heavy calls. Streaming is Universal Streaming ($0.15/hr base) or Universal-3.6 Pro Realtime ($0.45/hr list). You also get PII redaction of text and audio (beep or silence), diarization, sentiment, and EU residency.
- You pay for silence on both. AssemblyAI bills streaming by session duration. Deepgram bills per second of audio streamed, silent frames included. If the caller listens to TTS or pauses for 55% of a 6-minute call and you keep the audio flowing for barge-in, you're billed for all 6 minutes. That's 2.22x the list rate per spoken minute.
- 36-month TCO at 1.5M min/mo (900k streaming + 600k batch, HIPAA/PCI): at $75/hr senior CEE rates, an owned L4 GPU cluster with Silero VAD ($168,600) comes in $201,600 under AssemblyAI ($370,200, 4.25-month payback), $223,560 under Deepgram on Growth rates ($392,160, 3.88-month payback), and $388,800 under our OpenAI estimate ($557,400, 2.34-month payback). Below 150,000 min/mo the cloud APIs win at $700 to $1,100/mo.
2026 ASR internals & speech-to-text API pricing: Deepgram vs AssemblyAI vs OpenAI
A Deepgram vs AssemblyAI decision (or AssemblyAI vs Deepgram) comes down to four things that clean-read LibriSpeech scores don't show: turn-taking latency, accuracy on jargon and overlapping speakers, how silence is billed, and where the audio is allowed to live (HIPAA, VPC, EU).
1. Deepgram pricing and internals (Nova-3 & Flux): Keyterm Prompting & native EoT
Deepgram built its streaming models for low-latency AI agents. Three things matter here:
- Nova-3 pricing and Keyterm Prompting. On the Deepgram pricing page, pre-recorded Nova-3 is $0.0048/min ($0.288/hr) pay-as-you-go and $0.0042/min ($0.252/hr) on Growth, with multilingual at $0.0050 to $0.0058/min. Streaming is $0.0077/min ($0.462/hr) pay-as-you-go and $0.0065/min ($0.390/hr) on Growth.
nova-3-medicaltargets clinical speech (see the Deepgram Nova-3 docs). Deepgram Keyterm Prompting takes up to 500 tokens of keyterms per request (?keyterm=semaglutide&keyterm=SKU-9482-BX), so drug names, tickers, and part numbers get boosted at runtime with no retraining. - Flux and native End-of-Turn (EoT). A static silence timer (
endpointing=400ms) either cuts off a caller reading an account number or adds a sluggish pause after a short reply. Flux costs $0.0065/min pay-as-you-go and $0.0057/min on Growth, so it isn't priced above Nova-3. It makes the turn decision inside the model, from how the speech sounds and whether the sentence is complete, and emitsEndOfTurnat about 260 ms p50 by Deepgram's own figure. Seteager_eot_thresholdand you also getEagerEndOfTurn, which lets large language models start a reply early, andTurnResumedto cancel it if the caller keeps talking. - Self-hosted containers. Deepgram self-hosted containers run the models inside your AWS or GCP GPU VPC. They're sold on enterprise agreements.
2. AssemblyAI pricing and internals (Universal-2, Universal-3.5 Pro & Universal-3.6 Pro Realtime): crosstalk accuracy & Audio Intelligence
AssemblyAI aims its models at noisy telephony (8 kHz G.711 mulaw), heavy accents, and post-call compliance work:
- Universal-2 and Universal-3.5 Pro for batch. Among the AssemblyAI pre-recorded models, Universal-2 costs $0.15/hr ($0.0025/min), roughly 40 to 48% below Nova-3 batch depending on the Deepgram tier. Universal-3.5 Pro ($0.21/hr, $0.0035/min) is the production batch model AssemblyAI positions for overlapping speakers and alphanumeric strings. Those accuracy claims are the vendor's, so test on your own audio.
- Universal Streaming and Universal-3.6 Pro Realtime. AssemblyAI's streaming ASR starts at $0.15/hr ($0.0025/min) for Universal Streaming. Universal-3.6 Pro Realtime lists at $0.45/hr and adds semantic turn detection (
end_of_turn_confidence) and word boosting. We couldn't find a published volume price for it, so the $0.36/hr ($0.0060/min) in our TCO model is an assumed negotiated rate. - Audio Intelligence and EU residency. AssemblyAI PII Redaction covers 50+ entity types in text and can export MP3 or WAV files with beep or silence masking. On batch, text redaction is $0.08/hr and audio redaction is $0.05/hr. Speaker diarization and entity and sentiment detection are add-ons too, and the Dublin endpoint (
api.eu.assemblyai.com) covers EU residency for GDPR.
3. OpenAI (GPT-4o-transcribe / Realtime) vs Owned L4 GPU (Parakeet-TDT + Faster-Whisper)
On OpenAI audio pricing, gpt-4o-transcribe is about $0.006/min. Realtime audio input is billed per token ($32 per 1M audio input tokens for gpt-realtime), so the per-minute cost depends on your token rate. Diarization needs the separate gpt-4o-transcribe-diarize model, and there are no audio beep exports or VPC containers. Two open-weight engines make self-hosting practical on 24 GB AWS EC2 G6 (NVIDIA L4) instances:
nvidia/parakeet-tdt-0.6b-v2(CC-BY-4.0): a FastConformer Token-and-Duration Transducer (TDT) with about 6.05% average WER on the Open ASR leaderboard and more than 2,000x RTFx in NVIDIA's reported runs, with punctuation and word timestamps. Our build serves it through TensorRT.- SYSTRAN/faster-whisper (Large-v3-Turbo, MIT): a CTranslate2 float16 engine that covers 99 languages and runs several times faster than Large-v3.
4. Deepgram vs AssemblyAI streaming billing: why silence is charged (2.22x effective markup)
Both vendors charge you for silence, by different routes. AssemblyAI bills streaming by session duration, idle time included. Deepgram bills speech to text per second of audio you stream, and silent frames you keep sending are audio.
Take a 6-minute (360s) voice agent call. TTS playback is 36.7% of it, and LLM or CRM lookups plus thinking pauses are another 18.3%. That leaves 55% (198s) of non-speech and 45% (162s) of the caller actually talking. You can't just stop sending audio while the bot speaks, because the recognizer needs the caller's stream to catch a barge-in. Deepgram's KeepAlive messages hold the socket open without audio, but with no PCM flowing there's no live interruption detection either. So all 360 seconds are billed, and a $0.0065/min Growth rate becomes about $0.0144 per spoken minute (2.22x). Stereo SIP recording forks that stream the caller and agent legs separately double the charge again unless you mix or filter them upstream.
On an owned cluster, a CPU Silero VAD in your own gateway fixes this. It classifies 32ms frames, keeps 224ms of pre-roll so initial plosives (/p/, /t/, /k/) don't clip, and drops the 55% non-speech window before it reaches the GPU. Owning that gateway also loosens vendor lock-in, because the ASR engine behind it becomes swappable.
Cutting Deepgram vs AssemblyAI streaming cost with a Silero VAD silence gate
A production speech gateway with FinOps telemetry (see observability vs monitoring) has three failure modes to prevent: VAD state shared across callers, clipped first consonants when there's no 200 ms pre-roll buffer, and unredacted PHI or card numbers leaking into transcripts. That last one falls under the HIPAA Security Rule (45 CFR Part 164 Subpart C) and PCI DSS v4.0.1. For cards, keep the digits out of the audio path, for example as keypad tones masked at the SIP trunk (IETF RFC 4733 DTMF).
The part that saves money is small. Each caller gets its own Silero VAD state on ONNX Runtime, audio arrives as 16 kHz PCM16 in 32 ms frames, and only frames inside a speech window ever reach the GPU. A 7-frame ring buffer (224 ms) holds recent silence as pre-roll, so the first consonant is kept when speech starts. A speech probability of 0.52 or higher opens a turn, and 9 consecutive silent frames (288 ms) close it and hand the utterance to transcription and redaction. Silence outside a turn is dropped.
Around that gate sit a handful of design decisions. Each one maps to a cost or compliance risk from the sections above:
The cloud APIs force a different trade-off. With Deepgram or AssemblyAI you can't run this gate in front of the vendor without breaking barge-in, since their recognizers need the live caller stream. So the gate only pays off once you own the ASR behind it.
Deepgram vs AssemblyAI TCO: worked 36-month model at 1.5M minutes per month
At 1.5M minutes a month, an owned L4 cluster costs about $168,600 over 36 months, against $370,200 for AssemblyAI and $392,160 for Deepgram, under the assumptions below.
The model is a regulated voice agent and call-recording product over 36 months (54,000,000 minutes, 900,000 hours). Each month has 900,000 minutes (15,000 hours) of real-time WebSocket voice calls with a 55% silence and TTS-listening ratio, plus 600,000 minutes (10,000 hours) of async call recordings with speaker diarization (pyannote-audio or a cloud add-on), a HIPAA BAA, PCI DSS v4.0.1, and SOC 2 PII redaction.
Labor uses Idealogic's $75/hr senior CEE rate against a $165/hr US agency rate (2.2x). Both are our own working rates, not published statistics. The EU whole-economy hourly labour cost data (lc_lci_lev) and the O*NET Software Developers occupation (15-1252.00) show which way the wage gap runs, and neither gives an agency billing rate.
- Option A: Deepgram (Nova-3 streaming and batch on Growth rates, plus redaction and diarization). $12,000 setup (160h CEE at $75/hr; $26,400 US at $165/hr) + $353,160 API ($9,810/mo x 36: 900k streaming min at $0.0065 = $5,850 + 600k batch min at $0.0042 = $2,520 + $1,440 for PII redaction, diarization, and HIPAA add-ons, our estimate) + $27,000 ops (10h/mo x 36 = 360h CEE; $59,400 US) = $392,160 CEE ($438,960 US). Year-2+ run-rate: $126,720/yr CEE ($10,560/mo). On pay-as-you-go rates ($0.0077 and $0.0048) the API line is $11,250/mo.
- Option B: AssemblyAI (Universal-3.6 Pro Realtime + Universal-3.5 Pro batch + Audio Intelligence). $12,000 setup (160h CEE; $26,400 US) + $331,200 API ($9,200/mo x 36: 15,000 hrs of Universal-3.6 Pro Realtime at an assumed $0.36/hr volume rate = $5,400 + 10,000 hrs of Universal-3.5 Pro batch at $0.21/hr = $2,100 + $1,700 for PII audio redaction and diarization, our estimate) + $27,000 ops (10h/mo = 360h CEE; $59,400 US) = $370,200 CEE ($417,000 US). Year-2+ run-rate: $119,400/yr CEE ($9,950/mo). This prices AssemblyAI's premium streaming model. At the $0.45/hr list price the streaming line is $6,750/mo, and on Universal Streaming at $0.15/hr it drops to $2,250/mo.
- Option C: OpenAI (GPT-4o-transcribe + Realtime audio). $12,000 setup (160h CEE; $26,400 US) + $518,400 API ($14,400/mo x 36) + $27,000 ops (10h/mo = 360h CEE; $59,400 US) = $557,400 CEE ($604,200 US). Year-2+ run-rate: $181,800/yr CEE ($15,150/mo). The $14,400/mo is our estimate, since Realtime audio is billed per token, and it's the least certain line here.
- Option D: Owned L4 GPU ASR Cluster (NVIDIA Parakeet-TDT-0.6B-v2 + Faster-Whisper Large-v3-Turbo TensorRT/CTranslate2 + Silero VAD + Pyannote + Presidio on 4x AWS
g6.2xlarge1-yr RI): $39,000 CEE build (520h @ $75/hr; $85,800 US @ $165/hr) + $86,400 AWS GPU infra ($2,400/mo x 36: 4x RIg6.2xlargeL4 24GB + Spot batch burst + NLB/Redis/AWS KMS) + $43,200 CEE MLOps (16h/mo x 36 = 576h CEE; $95,040 US) = $168,600 CEE ($267,240 US). Year-2+ run-rate: $43,200/yr CEE ($3,600/mo).
CFO payback and the low-volume crossover
- Payback at 1.5M min/mo takes under 5 months. Option D costs $27,000 more to build ($39,000 minus $12,000) and runs at $3,600/mo. Against AssemblyAI the monthly gap is $6,350, so payback takes 4.25 months and the 36-month saving is $201,600. Against Deepgram it's $6,960/mo, 3.88 months, and $223,560. Against our OpenAI estimate it's $11,550/mo, 2.34 months, and $388,800. At $165/hr US labor the owned cluster ($267,240) still saves $149,760 to $336,960.
- Below 150,000 minutes a month, stay on the cloud APIs. Deepgram or AssemblyAI bills $700 to $1,100/mo at that volume. That beats $2,400/mo of baseline L4 GPUs plus 16h/mo of MLOps until ASR spend passes $4,000/mo or a hospital's VPC rules force in-VPC inference.
2026 ASR integration & owned L4 GPU build budgets (Senior CEE vs US)
Our AI software development, software development consulting, and dedicated development team squads deliver speech pipelines at $75/hr Senior CEE vs $165/hr US agency rates (2.2x):
Inside the worked 520-hour ($39,000 CEE vs $85,800 US) owned L4 GPU ASR build
- 1. TensorRT ASR Engine (30% / 156h / $11,700 CEE vs $25,740 US): Parakeet-TDT-0.6B-v2 (sub-120ms English turns) + Faster-Whisper Large-v3-Turbo (99-language fallback) workers on 4x
g6.2xlargeL4 nodes with OpenTelemetry GPU queue-depth autoscaling and dynamic batching. - 2. WebSocket + VAD Gate (25% / 130h / $9,750 CEE vs $21,450 US): Stateful Silero VAD pre-roll ring-buffer gateway (RFC 6455 WebSocket backpressure, Twilio Media Streams 8 kHz G.711 mulaw-to-16 kHz PCM16 polyphase resampler) dropping 55% non-speech duration before GPU dispatch.
- 3. Pyannote Diarization (20% / 104h / $7,800 CEE vs $17,160 US): Spot L4 speaker-turn segmentation, overlapping-crosstalk embedding clustering, and word-timestamp alignment driven by S3 event queues.
- 4. HIPAA PII Scrubber (15% / 78h / $5,850 CEE vs $12,870 US): Presidio NER + deterministic Luhn/SSN scrubber, FFmpeg timestamped audio beep/mute export to KMS-encrypted S3, and PG17 FORCE RLS append-only tables.
- 5. Keyterm & WER Eval (10% / 52h / $3,900 CEE vs $8,580 US): Runtime subword logit biasing dictionaries plus automated
jiwerCI regression gates across 250 noisy telephony customer fixtures.
Teams that argue over half a cent on the per-minute rate card tend to miss the bigger streaming cost: paying to transcribe more than three minutes of TTS playback and silence on every six-minute call. Under 150,000 monthly minutes, use Deepgram Flux for turn-taking and AssemblyAI Universal-3.5 Pro for noisy batch audio. Past that, put Silero VAD in front of four L4 GPUs before silence billing takes your voice AI gross margin.
Deepgram vs AssemblyAI: shipped Idealogic proof and a 5-step CTO checklist
Idealogic ships low-latency clinical, financial, and dispatch pipelines in production:
- Clinical & HIPAA privacy (AddMed / Pillbie): Encrypted patient boundaries and real-time adherence alerts in healthtech.
- Regulated financial ledgers (Fese & Chaindoc): Auditable PCI DSS v4.0.1 and cryptographic audit trails in fintech.
- Field dispatch telemetry (HaulBreeze & PikkUp): Real-time mobile sync over noisy cellular links in logistics and supply chain.
Five checks settle Deepgram vs AssemblyAI vs an owned L4 GPU cluster:
- Voice agent under 150k min/mo that needs fast turn-taking and drug or SKU jargon? Pick Deepgram Flux plus Nova-3 Keyterm Prompting.
- Async call-center QA, EU residency in Dublin, or redacted audio with beeps? Pick AssemblyAI Universal-2 ($0.15/hr) or Universal-3.5 Pro ($0.21/hr) with Audio Intelligence.
- Want a cloud hybrid? Run Deepgram Flux on live sockets and AssemblyAI Universal-3.5 Pro on post-call recordings. You'll manage two vendors.
- Taking card payments by phone? Keep the digits out of the transcription path, for example with keypad entry (RFC 4733 DTMF) masked at the SIP trunk.
- Passing 400k to 1.5M+ min/mo, or bound by strict VPC residency? Deploy owned L4 GPU ASR (Parakeet-TDT + Faster-Whisper + Silero VAD + Presidio). In our model that's $168.6k CEE over 36 months, $201.6k to $388.8k less than the cloud options. Contact our team for a 48-hour ASR TCO audit.

CTO and Co-founder at Idealogic. Directing engineering, systems architecture, security, and full-stack delivery across web, mobile, and blockchain.
Frequently asked questions
Pick Deepgram Flux and Nova-3 for conversational voice agents and real-time clinical scribes that need End-of-Turn detection inside the model (Deepgram reports about 260 ms p50), runtime keyterm prompting without retraining, or self-hosted containers in your own VPC on an enterprise agreement. Pick AssemblyAI when noisy multi-speaker call recordings matter most. Universal-2 and Universal-3.5 Pro cover batch, Universal-3.6 Pro Realtime covers streaming, PII redaction can export beeped or silenced audio, and there is an EU endpoint in Dublin.
On pre-recorded audio AssemblyAI is cheaper at list price. Universal-2 is $0.15 per hour ($0.0025 per minute) and Universal-3.5 Pro is $0.21 per hour ($0.0035 per minute). Deepgram Nova-3 pre-recorded is $0.0048 per minute pay-as-you-go and $0.0042 on Growth, with multilingual at $0.0050 to $0.0058. For streaming, AssemblyAI Universal Streaming starts at $0.15 per hour and Universal-3.6 Pro Realtime lists at $0.45 per hour. Deepgram Nova-3 streaming is $0.0077 per minute pay-as-you-go and $0.0065 on Growth, and Flux is $0.0065 or $0.0057. These are November 2026 list prices, and Deepgram marks some as promo pricing.
You pay for silence. AssemblyAI bills streaming by session duration, idle time included. Deepgram bills speech to text per second of audio streamed, and silent frames are still audio. A voice agent has to keep sending caller audio while the bot talks so barge-in works, so on a 6-minute call where the caller speaks 45% of the time you're billed for all 6 minutes. That's 2.22x the list rate per spoken minute, unless a Silero VAD gateway in front of your own ASR drops the non-speech frames first.
A fixed endpointing timer, say 500 ms of silence, either cuts off a caller who pauses mid-sentence or adds a sluggish gap after a quick yes or no. Deepgram Flux decides inside the recognition model, from how the speech sounds and whether the sentence is complete. It sends an EndOfTurn event at about 260 ms p50 by Deepgram's own figure. Set eager_eot_threshold and it also sends EagerEndOfTurn, so your LLM can start a reply early, and TurnResumed to cancel it when the caller keeps talking.
Older speech engines needed labeled audio and a paid fine-tuning run before they'd recognize rare drug names, tickers, or alphanumeric part numbers. Nova-3 takes keyterms as a request parameter, up to 500 tokens per request, so new vocabulary works on the next call with no retraining. AssemblyAI handles domain vocabulary in Universal-3.5 Pro and Universal-3.6 Pro Realtime through word boosting and entity formatting.
Both sign HIPAA Business Associate Agreements (Deepgram's is for Enterprise customers) and both hold SOC 2 Type II attestations. AssemblyAI runs an EU endpoint in Dublin. Deepgram documents self-hosted containers for your own AWS or GCP VPC, sold on enterprise agreements. Neither vendor takes card data out of PCI DSS v4.0.1 scope for you. Keep card digits out of the transcription path, for example by capturing them as DTMF keypad tones masked at the SIP trunk (RFC 4733), so they never reach a transcription socket or a transcript log.
Not below 150,000 minutes a month, where cloud APIs cost $700 to $1,100 a month and you run no GPUs. In our worked model at 1.5 million minutes a month (900,000 streaming, 600,000 batch), an owned cluster running NVIDIA Parakeet TDT and Faster Whisper Large v3 Turbo behind Silero VAD on four reserved AWS g6.2xlarge L4 instances totals $168,600 over 36 months at $75 per hour senior CEE rates. That's $201,600 less than AssemblyAI ($370,200) and $223,560 less than Deepgram on Growth rates ($392,160).
At Idealogic's senior Central and Eastern European rate of $75 per hour (versus $165 per hour at US agencies), a 2 to 3 week audit of STT latency, WER, and silence spend runs $11,250 to $18,000. A production Deepgram Flux or AssemblyAI Universal rollout with PII redaction runs $12,000 to $24,000. A 520-hour owned AWS L4 ASR build with TensorRT and CTranslate2, Silero VAD, Pyannote diarization, and Presidio scrubbing costs $39,000, or $85,800 at US rates.
More from the journal

Vapi vs Retell vs Bland AI: 2026 Voice AI Agent TCO
Engineering and 36-month TCO teardown of Vapi vs Retell, Vapi vs Bland AI, and ElevenLabs vs Deepgram in 2026: stacked per-minute pricing, concurrency caps, sub-500ms latency budgets, PCI DSS v4.0.1 DTMF guards, and owned LiveKit Agents.

LiveKit vs Agora vs Daily: 2026 WebRTC & Voice AI TCO
Engineering and 36-month TCO teardown of LiveKit vs Agora and Daily vs LiveKit in 2026: Apache 2.0 Go Pion SFU dynacast and LiveKit Agents vs Agora SD-RTN aggregate-resolution video tiers, HIPAA BAA Egress recording, and self-hosted AWS Graviton.

Portkey vs LiteLLM in 2026: AI Gateways, Latency & TCO
Engineering and 36-month TCO comparison of Portkey vs LiteLLM vs Helicone and compiled Go/Rust AI gateways in 2026: streaming overhead, enterprise SSO walls, PostgreSQL 17 FORCE RLS token-budget ledgers, and 120M req/mo math.