Now live New models for transcription and text-to-speech. AssemblyAI-STT-3.7 AssemblyAI-TTS-1
Insights & Use Cases

Speech recognition cost in 2026: total cost of ownership, API vs self-hosted

What speech recognition really costs in 2026: per-hour rates, API vs self-hosted GPU TCO, entity-error cost, voice-agent stack cost and volume spikes.

Abstract green cylinder illustration

Written by

Kelsey Foster

Published on

18 November 2026

Speech recognition pricing looks simple until you build something with it. One vendor quotes per minute, another per hour, a third meters audio tokens. Self-hosting promises a price of zero per hour, as long as you don't count the GPUs or the engineers. And the number on the rate card turns out to be a poor predictor of the number on the invoice.

This guide is about the invoice. It starts with today's rates, then builds the total cost of ownership (TCO) model most teams skip: API vs self-hosted GPU, the cost of errors, the full voice-agent stack, and what happens when volume spikes.

What it really costs to run speech-to-text (as of November 2026)

AssemblyAI's per-hour rates, as of November 17, 2026, from the pricing page:

Surface Model Rate Billed on
Pre-recorded (async) AssemblyAI-STT-3.7 $0.21/hr Audio duration, pro-rated to the second
Pre-recorded (async) Universal-2 (99 languages) $0.15/hr Audio duration, pro-rated to the second
Streaming (notetakers, captions, ambient) AssemblyAI-STT-3.7-Realtime $0.30/hr Session duration (how long the WebSocket stays open)
Streaming (voice agents) AssemblyAI-STT-3.7-Realtime-Agent $0.45/hr Session duration
Sync Universal-3.6 Pro $0.30/hr Audio sent
Text-to-speech AssemblyAI-TTS-1 $15 per 1M characters Input characters

There are no minimums, upfront fees or contracts on pay-as-you-go, and new accounts get $50 in free credits. For add-on rates and a multi-vendor comparison, see the speech-to-text API pricing guide. It holds the full rate card, so this post can focus on everything around it.

If you're on an older model: Universal-3.5 Pro for pre-recorded audio was deprecated on November 17, 2026, and from December 8, 2026, accounts that don't set speech_models move to AssemblyAI-STT-3.7 at the same $0.21/hr. Universal-Streaming English and Multilingual were deprecated the same day, with a tentative retirement date of February 1, 2027. Universal-3.6 Pro Realtime, the previous streaming model, is still fully supported at $0.45/hr.

Three things already move the bill before you've written any code:

  • Streaming bills session time, not audio time. Silence, LLM thinking time and the stretch where the caller listens to your agent all count. Close sessions when calls end; per the billing docs, an unterminated session auto-closes after 3 hours and bills for all of it.
  • Add-ons stack. Speaker diarization (+$0.02/hr async, +$0.12/hr streaming), prompting (+$0.05/hr), Voice Focus (+$0.10/hr on AssemblyAI-STT-3.7-Realtime-Agent) and PII Text Redaction (+$0.08/hr async, +$0.12/hr streaming) each add to the base rate. Keyterms are included on both 3.7 realtime models and on Universal-2, and cost +$0.05/hr on AssemblyAI-STT-3.7.
  • Multichannel bills per channel. A one-hour stereo file is billed as two hours.
Measure Your Real Cost, Not the Rate Card

Run a representative sample of your own audio on AssemblyAI-STT-3.7 and read the bill for your actual workload. $50 in free credits, no minimum, no contract.

Sign up free

TCO model: API vs self-hosted GPU

The pitch for self-hosting Whisper or Parakeet is that inference is free once you own the GPU. That pitch leaves out most of the bill.

What you're buying at $0.21/hr is 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, and the pager that doesn't go off at 2am when a batch job wedges. A self-hosted deployment that's production-grade, with monitoring, retries and a rollback plan, needs engineering time every month, not just at launch.

Here's a simple model you can rerun with your own numbers. Every input in it is an assumption, and the assumptions are generous to self-hosting.

Assumptions (illustrative, replace with your own):

  • A cloud GPU instance at $2.00/hr on demand, running 730 hours a month.
  • Each GPU sustains 960 audio hours of batch transcription per day (40x real time, with batching).
  • A floor of two GPUs, always on, for redundancy. Add one GPU per 960 audio hours per day, plus one spare.
  • Engineering at $200,000 a year loaded: 0.25 FTE to keep it healthy up to 10,000 hours a month, 0.5 FTE at 100,000.
  • The API comparison is AssemblyAI-STT-3.7 at $0.21/hr, pre-recorded, with no add-ons.
Monthly audio GPUs GPU cost Engineering Self-hosted total AssemblyAI-STT-3.7 API
1,000 hours 2 $2,920 $4,167 $7,087 $210
10,000 hours 2 $2,920 $4,167 $7,087 $2,100
100,000 hours 5 $7,300 $8,333 $15,633 $21,000

The break-even point comes from one formula: fixed self-hosted cost ÷ (API rate − GPU cost per audio hour). With these inputs, that's $7,087 ÷ ($0.21 − $0.05) ≈ 44,000 audio hours a month. Below that, the API is the smaller bill. Above it, self-hosting can come out ahead on infrastructure, if your throughput assumption holds.

Three things the table doesn't show, and all three push the break-even further out:

  • Streaming changes the math. Batch work can queue. Live calls can't, so self-hosted streaming has to be provisioned for peak concurrency, not average volume, and those GPUs sit idle off-peak.
  • Throughput is the swing variable. Halve the audio hours per GPU-day and the GPU cost per audio hour doubles. Benchmark your model on your hardware before you trust the spreadsheet.
  • Accuracy isn't free. If a self-hosted model makes more errors on your audio than the API, the review cost in the next section lands on the self-hosted side of the ledger.

Self-hosting makes sense when you have a hard data-residency constraint a managed API can't meet, when you're doing research rather than shipping a product, or when your volume is very large and your throughput is proven. For most teams, it converts a predictable operating expense into an unpredictable engineering one.

Effective cost: entity errors times review cost

Word error rate treats every word as equally expensive. "The" and "umm" and an account number all count once. But nobody reprocesses 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 trigger a human review, a callback or a support ticket, and each one dwarfs what you paid to transcribe the hour it came from. That's why word error rate is broken as a purchasing metric.

On AssemblyAI's English voice-agent benchmark (12,460 scenarios recorded as telephony audio with no, low and heavy background noise, November 2026), AssemblyAI-STT-3.7-Realtime-Agent has the lowest entity error rate of the 26 streaming models we ran, ahead of Soniox, ElevenLabs, Google, Speechmatics and Deepgram: 13.2%, against 15.2% for Soniox Realtime v6, 16.9% for ElevenLabs Scribe v3 Live and 25.1% for Deepgram Nova-3.

Now put money on it. Take a voice agent handling 200,000 three-minute calls a month (10,000 hours), with four entities per call, so 800,000 entities. Assume one in ten entity errors escalates to a human at $5 each. Both are illustrative inputs.

Streaming model Entity error rate Entity errors per month Escalation cost per month
AssemblyAI-STT-3.7-Realtime-Agent 13.2% 105,600 $52,800
Soniox Realtime v6 15.2% 121,600 $60,800
ElevenLabs Scribe v3 Live 16.9% 135,200 $67,600
Deepgram Nova-3 25.1% 200,800 $100,400

Now the speech-to-text line for the same agent. At 10,000 hours, AssemblyAI-STT-3.7-Realtime-Agent is $4,500 a month. Two of the vendors above list lower streaming rates: ElevenLabs lists Scribe v3 Live at $0.22/hr ($2,200 a month), and Deepgram lists Nova-3 Monolingual streaming at $0.0048/min ($0.288/hr, $2,880 a month) as promo pricing, against a $0.0077/min list rate, both on their own pricing pages as of November 17, 2026. In this example, each escalation line is more than ten times any speech-to-text line quoted here, and the gap between the top and bottom rows ($47,600 a month) is about 29 times the gap in speech-to-text spend ($1,620). Benchmark error rates won't match your calls exactly, so score entities on your own audio and plug in your real escalation cost. The full arithmetic is in the true cost of inaccurate transcription.

A clinical example

Healthcare is where error cost bites hardest. AssemblyAI-STT-3.7 at $0.21/hr plus Medical Mode at +$0.15/hr is $0.36/hr on pre-recorded audio. Medical Mode improves accuracy on medication names, generic and brand, and reaches a 3.2% missed entity rate, roughly 20% fewer missed medical entities than the base model, per the medical transcription use case page. It's turned on with domain: "medical-v1" and covers English, Spanish, German and French on pre-recorded and streaming audio. It isn't available on the Sync or Dictation APIs, so a clinical workload that needs it runs on async or streaming.

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.

Voice-agent stack TCO at 10,000 hours a month

This is the question that gets asked in budget meetings, and the rate card answers it worst. Ten thousand hours a month is a real production voice agent: about 200,000 three-minute conversations.

Cost line How it's metered Monthly cost at 10,000 hours
Speech-to-text AssemblyAI-STT-3.7-Realtime-Agent, $0.45/hr of session time $4,500
Text-to-speech AssemblyAI-TTS-1, $15 per 1M input characters About $3,240 if the agent talks 40% of the time at ~15 characters a second (216M characters)
LLM Per 1M input and output tokens through LLM Gateway, by model Varies with model, turn count, prompt length and context carried
Noise handling Voice Focus on AssemblyAI-STT-3.7-Realtime-Agent, +$0.10/hr of session time, optional $1,000 if enabled
Telephony Your carrier (for example, bring your own Twilio) Your carrier's rates
Orchestration LiveKit, Pipecat or your own loop, on your infrastructure Your hosting and on-call

Which line is biggest matters less than how each one is metered: duration-metered lines are forecastable and token-metered lines aren't. Ten thousand hours of AssemblyAI-STT-3.7-Realtime-Agent is $4,500 whether your system prompt is 200 tokens or 2,000. The LLM line moves every time you lengthen a prompt or carry more history, so budget it from a sample of real conversations, not a rate card. LLM Gateway prices on /pricing are for global routing; in-region US or EU endpoints are 10% higher.

Two things that look like free features are your code, not the vendor's. Turn detection is included in the streaming session, but barge-in isn't: the model sends a SpeechStarted message and your agent stops text-to-speech playback. And a streaming session left open after a call ends keeps billing. For the engineering-time side of the stack, see the hidden costs of a voice agent stack.

Where the Sync API changes the math

If your agent detects turns itself, you don't have to pay for session time at all. The Sync API takes one HTTP POST per finished utterance and returns a Universal-3.6 Pro transcript in the same call, at $0.30/hr of audio sent; /pricing lists results in about 134 ms at the median, for requests of up to 20 minutes. On a bursty, short-utterance workload, like push-to-talk, IVR or voicemail, paying for audio sent instead of connection time can be a large difference. Keyterms and conversation context are included; prompting is +$0.05/hr. The trade-off is feature surface: Medical Mode isn't available on Sync, and Sync doesn't run on the 3.7 models.

Size Your Effective Cost With an Expert

Get help modeling entity-error cost, volume planning and stack consolidation for your workload, with guidance tailored to your industry and use case.

Talk to AI expert

What happens when call volume spikes 10x overnight

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

On pay-as-you-go billing, your cost goes up 10x and nothing else happens. That sounds obvious, but on many enterprise contracts a 10x spike triggers something worse than a 10x bill. Committed-volume contracts have overage rates. Plan-tier products cap concurrency, so you don't overspend, you fail: connections get refused while callers are on the line. Per-seat pricing means scaling waits on procurement.

AssemblyAI has no committed tier to exceed and no concurrency fee. Streaming limits the number of new sessions per minute, with no cap on open sessions. Per the streaming rate limits docs, paid accounts start at 200 new sessions per minute, and whenever you open 80% or more of your limit in a minute, the next minute's limit rises by 10%. Your bill tracks usage in both directions.

Reliability under a spike is a cost line even when it doesn't appear on the invoice: a refused connection on a live call is a customer complaint, and a dropped session in a notetaker is a meeting with no notes.

When switching providers pays for itself

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

  • Your review queue is a line item. If humans correct transcripts as a standing process, the accuracy gap already costs more than any rate difference.
  • Entities fail on your hardest audio. Pull 100 real calls with accents, crosstalk and account numbers, and count the entity errors.
  • You're paying for features you don't use. À la carte add-ons let you turn on Entity Detection at $0.08/hr without buying a bundle.
  • Your bill doesn't track your usage. Per-minute rounding, minimum commitments and concurrency fees decouple cost from consumption.
  • You're running four vendors to make one voice agent work. Consolidation pays off in integration surface, incident triage and contract count, as well as rate.

Match the model to the job before you compare rates. A notetaker or live-caption product doesn't need the voice-agent model: AssemblyAI-STT-3.7-Realtime at $0.30/hr is built for that work, with realtime speaker labels. AssemblyAI-STT-3.7-Realtime-Agent at $0.45/hr is for audio an agent has to answer.

Rough volume guidance, as planning assumptions: under 500 hours a month, stay pay-as-you-go and don't spend a week modeling it. Between 500 and 2,000, model committed volume against what flexibility is worth to you. Above 2,000, it's worth a pricing conversation, but go in knowing your effective cost, not just your rate. If accuracy is the reason you're switching, the real-time speech-to-text API comparison ranks the field on published benchmarks.

Where this is going

Speech recognition rates will keep falling. That's what happens to inference. The more interesting change is 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, recordings and the infrastructure underneath. The per-hour transcription rate is one line in that bill, and often not the biggest one. The right question is what a completed unit of work costs you, end to end, including the errors you fix and the infrastructure you run. Most teams still can't answer that about their current stack, and the ones who can make better procurement decisions.

See Voice AI in Action

Test streaming and pre-recorded transcription on your own audio, then price the add-ons you need before you commit to anything.

Try playground

Frequently Asked Questions (FAQs)

How much does speech recognition cost?

As of November 17, 2026, AssemblyAI's speech recognition costs $0.21/hr for pre-recorded audio on AssemblyAI-STT-3.7, $0.30/hr for live transcription on AssemblyAI-STT-3.7-Realtime, and $0.45/hr for voice agents on AssemblyAI-STT-3.7-Realtime-Agent. Add-ons such as speaker diarization or Medical Mode bill on top. Human transcription, by comparison, is usually priced per minute at a far higher rate, and errors you correct by hand add to the real cost of either.

What's the total cost of ownership of self-hosting Whisper?

Self-hosting Whisper costs GPU time plus the engineering to run it, and the engineering usually dominates at low volume. With generous assumptions ($2.00/hr GPUs, a two-GPU floor and a quarter of an engineer), a production deployment runs about $7,000 a month before it transcribes much audio. A managed API at $0.21/hr bills $2,100 for 10,000 hours.

At what volume does self-hosted speech recognition break even with an API?

Break-even is your fixed self-hosted cost divided by the gap between the API rate and your GPU cost per audio hour. With illustrative inputs ($7,087 a month fixed, $0.05 GPU cost per audio hour, $0.21/hr API), it lands around 44,000 audio hours a month. Lower GPU throughput, streaming workloads or extra review cost push it further out.

How do speech recognition errors affect total cost?

Entity errors, such as wrong account numbers or misspelled names, trigger human review, callbacks and escalations that cost far more than transcription. On a 10,000-hour voice agent with four entities per call, the gap between a 13.2% and a 25.1% entity error rate is about 95,000 extra errors a month. Price errors at your real escalation cost before you compare per-hour rates.

Is streaming speech-to-text billed on audio time or session time?

AssemblyAI streaming is billed on session time: how long the WebSocket stays open, not how much audio you send. Idle time between turns counts, and an unterminated session auto-closes after 3 hours and bills for the full duration. The Sync API, by contrast, bills on the audio you send, at $0.30/hr.

Do speech-to-text add-ons stack on top of the base rate?

Yes. Each add-on bills per hour on top of the base model rate, only when you turn it on. For example, AssemblyAI-STT-3.7-Realtime-Agent with streaming diarization (+$0.12/hr) and Voice Focus (+$0.10/hr) is $0.67/hr of session time. Keyterms are included on both 3.7 realtime models.