Optimizing Voice AI costs: When to switch STT providers and what to expect

Speech recognition pricing looks simple until you build something with it. One vendor quotes per minute, another per hour, a third meters audio tokens. Add-ons appear as line items or don't appear at all. And the number on the rate card turns out to be a poor predictor of the number on the invoice.

So let's start with the rate card anyway — because you can't reason about effective cost until you know the sticker price — and then work outward to the parts that actually move your bill.

What speech recognition costs in 2026

AssemblyAI bills four distinct surfaces. Each one solves a different problem, and each one has a different rate.

There are no minimums, no upfront commitments, and no contract required to start. You can sign up, hit the API, and be billed for exactly what you used. Concurrency is unlimited — there's no per-stream fee and no plan tier that caps how many conversations you can run at once.

And the price doesn't change by region. EU data residency runs on api.eu.assemblyai.com and streaming.eu.assemblyai.com at the same rate as US endpoints. Data residency is a deployment choice here, not an upsell.

The full itemized rate card, including every add-on, lives on the pricing page.

The free tier, stated properly

AssemblyAI's free tier is 185 hours of pre-recorded transcription and 333 hours of streaming. Those are separate allowances, not one pool — which matters, because a team prototyping a voice agent burns streaming hours and a team indexing an archive burns pre-recorded hours, and neither one eats into the other.

For context, here's what the hyperscalers give you:

Sixty minutes a month is a demo budget. It is not enough audio to find out whether a model handles your accents, your background noise, or your account numbers — which is the only question a free tier exists to answer. Two hundred calls at three minutes each is ten hours, and you're already out of runway on three of these four.

Evaluate on your own audio, at a volume large enough to be representative. Everything you learn from a clean read-speech sample, you will unlearn in production.

Add-ons: the line items that actually move your bill

Base rates are the easy part. Add-ons are where cost models go wrong, usually because someone budgeted the base rate and then turned on four features.

Here's the honest version. Note that several things people expect to pay for are included.

Noisy rooms, speakerphones, drive-time calls. Near-field and far-field profiles

Two of those deserve a second look. Keyterms prompting is included on the streaming flagship and on Universal-2 — competitors routinely meter custom vocabulary as a premium feature. And streaming diarization at +$0.12/hr buys live speaker labels that re-cluster at stream end, sending one correction within about half a second, for up to ten speakers.

Speech Understanding features are priced à la carte per hour on top of transcription: Translation $0.06, Entity Detection $0.08, Topic Detection $0.15, Content Moderation $0.15, PII text redaction $0.08, PII audio redaction $0.05, Sentiment $0.02, Key Phrases $0.01, Custom Formatting $0.03, Profanity $0.01, Speaker Identification $0.02, and Summarization $0.03. You pay for the ones you call. Nothing is bundled into a plan you have to buy wholesale.

The Sync API, and why most cost models miss it

Between "upload a file and poll for a result" and "hold a WebSocket open for the length of a conversation," there's a third shape: one HTTP POST in, a finished transcript back in the same response.

That's the Sync Speech-to-Text API — POST https://sync.assemblyai.com/transcribe with an X-AAI-Model: universal-3-5-pro header — at $0.45/hr. It returns in about 134 ms at p50 on a two-second clip, against five to six seconds for a submit-and-poll round trip on the async path. Clips run from 80 ms to two minutes, up to 40 MB.

The cost argument for Sync isn't the rate. It's what the rate replaces. If you're building dictation, push-to-talk, IVR, voicemail, or a voice agent that handles its own turn detection, the streaming path bills you for session duration — including the silence, the thinking, and the time the caller spends listening to your agent talk. Sync bills you for the audio you actually sent. On a bursty, short-utterance workload, that difference is not small.

The tradeoff is feature surface. Sync does not support PII redaction, speaker diarization, Speech Understanding, or Medical Mode. Keyterms and conversation context are included. If your workload needs redaction or speaker labels, you're on async or streaming — and that's a cost input, not a footnote. The Sync quickstart has the full parameter surface.

Why missed entities, not word error rate, set your effective cost

Here's the thing about word error rate: it treats every word as equally expensive. "The" and "umm" and "Levaquin" all count once. But nobody has ever reprocessed a transcript because the model dropped an article.

What costs money is a missed entity. A wrong account number. A mangled drug name. A phone number off by a digit. Those are the errors that trigger a human review, a callback, a re-run, or a support ticket — and the cost of each one dwarfs the $0.21 you paid to transcribe the hour it came from. This is why we've argued that word error rate is broken as a purchasing metric, and why entity accuracy is the number to shop on.

On Pipecat's open speech-to-text benchmark, which uses real agent conversations rather than clean read speech, Universal-3.5 Pro Realtime posts a 15.31% entity error rate — names 16.92%, places 6.28%, phone numbers 3.55%. Deepgram Flux is at 50.50% on entities. ElevenLabs Scribe v2 is at 39.70%. Google Chirp3 is at 21.51%, within a couple of points of us on pooled word error rate and roughly 40% worse on the entities. The benchmark is open source and reproducible.

Word error rate as supporting context, since you'll want it: 6.99% pooled on that same streaming benchmark against Deepgram Flux at 15.58%, ElevenLabs Scribe v2 at 9.76%, and Google Chirp3 at 9.04%. On the pre-recorded side, our async benchmark suite — 250+ hours, 80,000+ files, 26 datasets — puts English mean word error rate at 5.6% and median at 4.9%. On the Hugging Face Open ASR Leaderboard we sit at 5.03 average word error rate. Full methodology is on the benchmarks hub, and if you want to run your own comparison, we've written up how to evaluate speech recognition models without fooling yourself.

Translate that into money. A transcript at 95% accuracy needs roughly five minutes of human review per hour of audio. Drop to 85% and you're looking at 15 to 20 minutes. At a $0.50-per-minute reviewer cost, that gap is $5 to $7.50 per hour of audio — twenty to thirty times the base transcription rate. The cheaper API is frequently the more expensive decision, and it takes about one quarter of production data to find that out. We've laid out the full arithmetic in the true cost of inaccurate transcription.

One clinical example, because healthcare is where this bites hardest. Universal-3.5 Pro at $0.21/hr plus Medical Mode at $0.15/hr is $0.36/hr all-in on pre-recorded audio, with a 3.2% missed entity rate on medical terms — roughly 20% fewer missed medical entities than the base model. Medical Mode covers English, Spanish, German, and French, on pre-recorded and streaming. It is not available on the Sync API, so a clinical workload that needs it lives on async or streaming.

On healthcare specifically, one wording note that matters: AssemblyAI enables covered entities and their business associates subject to HIPAA to use the AssemblyAI services to process protected health information (PHI). AssemblyAI is considered a business associate under HIPAA, and we offer a standard Business Associate Addendum (BAA) that is required under HIPAA to ensure that AssemblyAI appropriately safeguards PHI. The BAA can be signed in minutes without a sales call.

Total cost of ownership for a voice agent at 10,000 hours a month

This is the question that actually gets asked in budget meetings, and it's the one the rate card answers worst.

Ten thousand hours a month is a real production voice agent — call it 200,000 three-minute conversations. On the Voice Agent API at a flat $4.50/hr, that's $45,000 a month. One number, one invoice, one vendor. Every feature is included in the $4.50/hr rate. There are no per-layer add-ons, concurrency fees, or per-agent subscriptions.

Now price the same agent assembled from parts.

The point isn't that assembled stacks are always more expensive per hour. Sometimes they aren't — if you already run inference infrastructure and your conversations are short and scripted, you can build something cheaper. The point is that a flat rate is forecastable and a metered stack is not.

Token-metered pricing is the sharpest version of this. OpenAI's Realtime API doesn't publish an hourly rate at all; it publishes audio token prices — $32.00 per 1M input tokens and $64.00 per 1M output tokens on gpt-realtime-2.1. What that costs per hour depends on how much your caller talks, how much your agent talks back, how long your system prompt is, and how much conversation history you re-send each turn. All four of those things change when you improve the agent. Your unit economics move every time you ship.

At $4.50/hr, they don't. A finance team can multiply. That's the whole argument, and we've written more on the hidden costs of a voice agent stack if you want the engineering-time side of it too.

What happens when call volume spikes 10x overnight

A campaign lands. A competitor goes down. A product goes viral. Monday you were running 1,000 hours a month; Thursday you're running 10,000.

On per-second, pay-as-you-go billing, your cost goes up 10x and nothing else happens. That sounds obvious. It isn't, because on most enterprise contracts a 10x spike triggers something worse than a 10x bill.

Committed-volume contracts have an overage rate, and the overage rate is usually worse than the rate you negotiated — that's the mechanism that makes the commitment worth signing in the first place. Plan-tier products cap concurrency, so you don't overspend, you fail: connections get refused while your callers are on the line. Per-agent and per-seat pricing means scaling means a procurement cycle, and procurement cycles do not run at the speed of a traffic spike.

Here's what doesn't happen on AssemblyAI. There's no committed tier to exceed, so there's no overage rate. Concurrency is unlimited, and the platform runs 100,000 concurrent streams and 600M+ inference calls a month at 99.99% uptime. There's no plan to upgrade and no approval to wait on. Your bill is linear in your usage, in both directions — the spike costs 10x, and the month after the spike costs what that month's traffic costs.

The asymmetry worth naming: a committed contract punishes you for guessing low and for guessing high. Pay-as-you-go asks you not to guess.

"The transcription accuracy, reliability, and speed of AssemblyAI's API have greatly enhanced our operations, reinforcing our trust in their technology and solidifying our partnership."

— Raj Shankar, SVP Product, Calabrio

Calabrio runs contact center workloads where volume is genuinely unpredictable and a dropped connection is a customer complaint. They've reported an 80% increase in customer satisfaction since the integration. Reliability at scale is a cost line even when it doesn't appear on the invoice — it appears on someone else's, in the form of a support queue.

Self-hosted Whisper: the cost and accuracy tradeoff

The pitch for self-hosting Whisper is that inference is free once you own the GPU. The pitch is not wrong. It's just incomplete.

What you're actually buying at $0.21/hr is not the forward pass. It's the GPU fleet you don't provision, the autoscaling you don't write, the queue you don't operate, the model upgrades you don't schedule, the multi-region failover you don't build, and the pager that doesn't go off at 2am when a batch job wedges. A first integration against a managed API typically runs 40 to 80 engineering hours. A self-hosted deployment that's actually production-grade — with monitoring, retries, and a rollback plan — is a different order of magnitude, and then it needs 5 to 10 hours a month forever.

Run the break-even honestly. At $0.21/hr, 10,000 hours a month is $2,100. If self-hosting takes one engineer even a quarter of their time to keep healthy, you are not saving money at that volume. The math flips somewhere north of it, and where exactly depends on your loaded engineering cost — but the crossover is much further out than the "inference is free" framing suggests.

Then there's accuracy. Universal-3.5 Pro has a hallucination rate roughly 30% lower than Whisper — and hallucinations are the expensive failure mode, because unlike a garbled word, a fluent invented sentence doesn't look wrong. It gets past review. It lands in a summary, a CRM field, or a clinical note. You find it later, from a customer.

Self-hosting makes sense when you have a hard data-residency constraint that a managed API can't meet, when you're running research rather than production, or when your volume is enormous and your accuracy tolerance is loose. For most teams shipping a product, it's a way to convert a predictable operating expense into an unpredictable engineering one.

When switching providers pays for itself

Switching costs real time, so the bar should be high. These are the five signals that clear it.

Your review queue is a line item. If humans are correcting transcripts as a standing process, the accuracy gap is already costing more than the rate difference. Measure the review hours before you measure the rate card.

Entities fail on your hardest audio. Not your demo audio. Pull 100 real calls with accents, crosstalk, and account numbers, and count the entity errors. That's your actual accuracy.

You're paying for features you don't use. Bundled plans charge for the whole shelf. À la carte pricing lets you turn on Entity Detection at $0.08/hr without buying a Speech Understanding tier.

Your bill doesn't track your usage. Per-minute rounding, minimum commitments, and concurrency fees all decouple cost from consumption. If you can't predict next month's invoice from next month's traffic, the pricing model is the problem.

You're running four vendors to make one voice agent work. Consolidation is worth real money — not primarily in rate, but in integration surface, incident triage, and contract count.

Rough volume guidance, and treat these as planning assumptions rather than laws: under 500 hours a month, stay pay-as-you-go and don't spend a week modeling it. Between 500 and 2,000, start modeling committed volume against what the flexibility is worth to you. Above 2,000, it's worth a pricing conversation — but go in knowing what your effective cost is, not just your rate. For a broader comparison across the market, we keep a running breakdown of speech-to-text API pricing.

Where this is going

The interesting thing about speech recognition pricing in 2026 isn't that rates are falling. They are, and they'll keep falling — that's what happens to inference.

It's that the unit of pricing is shifting. Five years ago you bought minutes of transcription. Today you're buying a conversation: speech in, reasoning, speech out, turn-taking, interruption handling, recordings, and the infrastructure underneath it, quoted as one number. The $4.50/hr flat rate isn't a discount on a bundle of parts. It's a different product with a different unit.

Which means the comparison that matters is changing too. Comparing per-hour transcription rates across vendors is becoming the wrong question — like comparing CPU prices when what you're buying is a database. The right question is what a completed unit of work costs you, end to end, including the errors you have to fix and the infrastructure you have to run. Most teams still can't answer that about their current stack. The ones who can are the ones making good procurement decisions.

Frequently asked questions

How much does AssemblyAI cost?

AssemblyAI costs $0.21/hr for pre-recorded transcription with Universal-3.5 Pro, $0.15/hr with Universal-2, $0.45/hr for streaming with Universal-3.5 Pro Realtime, $0.45/hr for the Sync API, and a flat $4.50/hr for the Voice Agent API. Everything is billed per second of usage with no minimums and no contract. The free tier is 185 hours of pre-recorded transcription plus 333 hours of streaming. Add-ons like Medical Mode (+$0.15/hr) and prompting (+$0.05/hr) are priced separately and only bill when you use them.

What's the total cost of ownership for a voice agent handling 10,000 hours per month?

On AssemblyAI's Voice Agent API, 10,000 hours a month is $45,000 — a flat $4.50/hr with speech-to-text, LLM inference, text-to-speech, turn detection, interruption handling, Voice Focus, recordings, transcripts, and infrastructure hosting all included. There are no per-layer add-ons, concurrency fees, or per-agent subscriptions, and bring-your-own Twilio SIP trunking carries no per-minute markup. An assembled stack at the same volume meters nine separate lines, each of which changes when you tune the agent, so the total is variable rather than forecastable.

How does flat-rate voice agent pricing compare to per-component STT plus LLM plus TTS billing at scale?

Flat-rate pricing is forecastable; per-component billing is not, because three of the components meter on things you change every time you ship. Speech-to-text bills on audio or session duration, text-to-speech bills per character or second, and LLM inference bills per 1M input and output tokens — so lengthening a system prompt or carrying more conversation history moves your unit cost. At $4.50/hr flat, 10,000 hours is $45,000 regardless of turn count, prompt length, or how much context the agent carries. Assembled stacks can win on raw rate for short, scripted conversations; they rarely win on predictability.

What happens to my costs if my voice agent call volume spikes 10x overnight?

Your bill goes up 10x and nothing else changes — there's no overage rate, no concurrency cap, and no plan upgrade to request, because AssemblyAI has no minimum commitment to exceed. Billing is per second and pay-as-you-go, so cost is linear in usage in both directions: the spike costs 10x, and the following month costs whatever that month's traffic costs. Concurrency is unlimited, and the platform runs 100,000 concurrent streams and 600M+ inference calls a month at 99.99% uptime. On a committed-volume contract, the same spike typically triggers overage rates worse than your negotiated rate, or refused connections at a concurrency ceiling.

Voice agent API pricing comparison: which is most cost-effective per hour at scale?

At scale, the most cost-effective option is the one with the fewest metered lines, because every metered line is a variable you have to forecast. AssemblyAI's Voice Agent API is a flat $4.50/hr ($0.075/min) billed per second of connected conversation, with every feature included and no concurrency or per-agent fees. Token-metered alternatives publish per-token rates rather than hourly ones — OpenAI's Realtime API lists $32.00 per 1M audio input tokens and $64.00 per 1M audio output tokens — which means the effective hourly cost depends on turn length, prompt size, and context carried. Build your comparison on a representative sample of your own conversations, not on a rate card.

Does AssemblyAI require a minimum commitment or contract to start billing?

No. There is no minimum commitment, no upfront spend, no contract, and no sales call required to start — you sign up, get 185 hours of pre-recorded and 333 hours of streaming free, and pay per second for what you use after that. Concurrency is unlimited with no per-stream fee, and EU data residency on api.eu.assemblyai.com costs the same as US endpoints. For teams processing PHI, the standard Business Associate Addendum can be signed in minutes without a sales call. Custom pricing conversations are available for high-volume workloads, but they're optional rather than a gate.