Optimizing Voice AI costs: When to switch STT providers and what to expect
Speech-to-text pricing runs about $0.15–$0.45 per hour, but the number that sets your bill is effective cost—what you pay after accuracy, rework, and add-ons.
Here's the short version, because it's the question everyone actually asks first: speech-to-text pricing usually runs from about $0.15 to $0.45 per hour of audio with a modern API, billed per second with no minimums. AssemblyAI's async flagship, Universal-3.5 Pro, is $0.21/hr; streaming with Universal-3.5 Pro Realtime is $0.45/hr base; and the value tier, Universal-2, is $0.15/hr. But the sticker rate is the least interesting number here. The one that decides your bill is your effective cost — what you pay after accuracy, rework, engineering time, and feature add-ons are all in the picture.
So let's do the honest version of this. Not "our per-minute rate is lower than theirs," but how speech-to-text pricing actually works, where the hidden costs hide, and when it's genuinely worth switching providers versus staying put.
How speech-to-text pricing works
Most speech recognition providers price on one of a few models. Understanding which one you're on — and which one you should be on — is the first place teams leave money on the table.
Per-minute vs per-hour billing
Some providers quote per minute, some per hour, and the framing matters more than you'd think. A rate of "$0.0075 per minute" sounds tiny until you multiply it out to $0.45/hr and compare it against an hourly quote. Worse, some vendors round every request up to the nearest 15 or 60 seconds, so a stream of short 8-second utterances — exactly what a voice agent produces — gets billed as if each one were much longer.
The thing to look for is per-second billing with no rounding and no minimums. AssemblyAI bills per second of audio processed, so 8 seconds costs you 8 seconds. For high-volume, short-utterance workloads that difference alone can move your bill 20–30% before you've changed a single line of code.
Usage-based vs committed volume pricing
Pay-as-you-go means you pay for what you use, month to month, with no floor. Committed-volume (or "reserved") pricing gives you a discount in exchange for a spend commitment — annual contracts, minimum monthly floors, or prepaid credits.
Committed pricing looks cheaper on the rate card. It isn't always cheaper in practice. If your volume is seasonal, still ramping, or hard to forecast, a commitment you don't fully consume is just a more expensive way to buy the hours you actually used. As a rough rule: stay pay-as-you-go until your monthly volume is both large and predictable, then negotiate.
Free tiers and when you'll hit paid usage
Free tiers are for evaluation, not production. They're small on purpose:
- Google Cloud Speech-to-Text: 60 minutes per month, ongoing.
- AWS Transcribe: 60 minutes per month, first 12 months only.
- Azure Speech: 300 minutes (5 hours) per month.
AssemblyAI gives new accounts free API credit to start — no credit card required — plus an open playground so you can test accuracy on your own audio before you spend anything. If you want a deeper tour of what's genuinely usable for free, we keep a running guide to the top free speech-to-text APIs and open-source engines. The point is the same everywhere: a few hundred minutes disappears fast, and the real question is what per-hour pricing looks like once you're past the free tier.
The total cost beyond the per-hour rate
This is the part most pricing comparisons skip, and it's where the actual money is.
How accuracy quietly sets your effective cost
A lower per-hour rate on a less accurate model is one of the most expensive bargains in this category. Every error is either a human minute spent correcting it or a downstream mistake — a misheard order, a wrong phone number, a broken voice-agent turn.
Put rough numbers on it. A transcript at 95% accuracy needs roughly 5 minutes of human review per hour of audio. Drop to 85% and you're looking at 15–20 minutes per hour. If a reviewer costs even $0.50/minute, that gap is $5–7.50 per hour of audio — an order of magnitude larger than any per-hour rate difference between providers. The cheap model just got very expensive.
Accuracy differences are real and measurable. In our own English testing, the previous-generation Universal-3 Pro landed at a 5.6% mean word error rate versus Universal-2's 6.1% — and on a harder, normalized code-switching benchmark, Universal-3.5 Pro posts a 7.69 average WER, ahead of ElevenLabs Scribe v2 (8.77) and Deepgram Nova-3 (12.22). A couple of points of WER doesn't sound like much on a slide. On your rework line, it's the whole story. If you want the methodology, here's how accurate speech-to-text really is and how to evaluate speech recognition models on your own data.
Infrastructure and integration costs
The API rate is one line item. The others:
- Engineering time. A first integration typically runs 40–80 hours. A provider with a clean SDK, clear docs, and no boilerplate WebSocket wrangling can cut that meaningfully.
- Ongoing maintenance. Budget 5–10 hours a month for monitoring, error handling, and SDK upgrades.
- Storage and egress. Some providers — Google and AWS among them — require your audio to already sit in their cloud storage (a GCS bucket or an S3 bucket) before you can transcribe it, which quietly pulls you toward their broader ecosystem and its bills.
Self-hosting an open-source model looks free until you price the GPUs, the MLOps headcount, and the accuracy gap. For most teams the build-vs-buy math lands on "buy," but it's worth running honestly.
The cost-vs-quality tradeoff, calculated
The formula that matters isn't rate per hour. It's:
Effective cost = API rate + (error rate × cost to correct each error) + amortized engineering + feature add-ons
Run that once with real numbers for your workload and the "expensive" high-accuracy provider is very often the cheapest total. That's the entire case for looking past the rate card.
Consolidating the stack: the Voice Agent API at a flat $4.50/hr
Here's where the pricing conversation has genuinely changed, and it's the cleanest cost story we have.
If you're building a voice agent, you're normally paying for three separate meters: speech-to-text, an LLM, and text-to-speech — each with its own vendor, its own rate card, its own rounding rules. The bill is unpredictable because it's the sum of three unpredictable things, and you're maintaining three integrations to get there.
The Voice Agent API collapses that into one WebSocket at a flat $4.50/hr, billed per second of session — STT, LLM, and TTS included. One meter, one integration, one predictable number.
Two things make that a switch-worthy line item. First, predictability: you can forecast a flat hourly rate, which you can't do cleanly with three metered vendors stacked on top of each other. Second, the comparison: OpenAI's Realtime API runs roughly $18/hr for a comparable speech-to-speech setup — the Voice Agent API is about 4x cheaper, with dedicated best-in-class models at each step and neural end-of-turn detection instead of a plain silence timer. If your STT-plus-LLM-plus-TTS stack is drifting past $4.50/hr all-in — and most do once you count the integration tax — consolidation is the move. Full breakdown lives on the Voice Agent API product page and the pricing page.
Stat callout: Voice Agent API — flat $4.50/hr, all-in (STT + LLM + TTS), one WebSocket. Roughly 4x cheaper than OpenAI Realtime's ~$18/hr, and one bill instead of three.
Right-sizing features so you don't overpay
Modern speech-to-text APIs are modular. You add capabilities as line items, which is great for cost control — as long as you're only paying for what you actually use. AssemblyAI's add-ons stack on top of the base transcription rate:
Noisy or multi-source audio that needs cleanup
Take a concrete example. Say you're transcribing clinical conversations. Your effective rate is Universal-3.5 Pro at $0.21/hr plus Medical Mode at $0.15/hr — $0.36/hr all-in, with a model tuned to catch the medical terms a general model drops. Add diarization at $0.02/hr if you need speaker separation and you're at $0.38/hr. That's the whole bill: transparent, itemized, and only for what the use case demands. Medical Mode supports English, Spanish, German, and French.
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 Business Associate Addendum (BAA) that is required under HIPAA to ensure that AssemblyAI appropriately safeguards PHI.
The mistake to avoid runs the other way too: paying for diarization, isolation, and keyterms on every job when your workload is clean single-speaker audio that needs none of it. Turn the extras off where you don't need them.
Try it on your own audio. The fastest way to size your real cost is to run a representative sample and read the itemized rate. Get your free API key — free credit to start, no credit card.
When to switch speech recognition providers
Switching has a real cost — re-integration, re-testing, retraining your team — so it should clear a bar. These are the situations where it usually does:
- Your rework bill dwarfs your API bill. If humans are spending real time fixing transcripts, a more accurate model pays for itself immediately, even at a higher per-hour rate.
- Your provider can't do what you now need. Real-time streaming, speaker diarization, multilingual and code-switching support, or a native voice-agent path — if you're bolting these on with duct tape, a provider that ships them natively is cheaper in total.
- You're running a three-vendor voice stack. Separate STT, LLM, and TTS bills that add up past $4.50/hr are a consolidation opportunity, both in dollars and in maintenance.
- Pricing is unpredictable. Aggressive rounding, surprise minimums, or opaque tiers make forecasting impossible. Flat, per-second, pay-as-you-go pricing fixes that.
- You've outgrown the free tier and the paid rates surprised you. The jump from free to production is where a lot of "cheap" providers stop being cheap.
If none of these apply, don't switch. Migration for a rounding-error rate difference isn't worth the engineering.
Cost optimization strategies that actually work
Volume discount breakpoints
Discount tiers cluster around predictable breakpoints. As a general shape:
- Under 500 hours/month: stay pay-as-you-go. A commitment won't beat per-second pricing at this scale.
- 500–2,000 hours/month: start modeling committed volume, but only if your usage is steady month to month.
- 2,000+ hours/month: negotiate custom pricing. At this scale the discount is worth a conversation.
Strategic volume planning
If you're near a breakpoint, it's sometimes worth consolidating workloads onto one provider to cross a discount threshold rather than splitting volume across two and qualifying for neither. But weigh that against concentration risk and the value of a fallback. AssemblyAI supports an ordered fallback list — latest model first, stable model as backup — so you can chase accuracy without giving up reliability.
Match the model to the job
Not every workload needs the flagship. Universal-2 at $0.15/hr covers 99+ languages and is a strong value tier for high-volume, cost-sensitive jobs; Universal-3.5 Pro at $0.21/hr is where you go for the best accuracy, diarization, and code-switching. Using a fallback list — ["universal-3-5-pro", "universal-2"] — lets you default to the flagship and fall back gracefully. And for summaries or chapters, you no longer pay for a separate feature: transcribe, then send the text through the LLM Gateway to the model of your choice.
Looking ahead
The direction of travel is clear: per-hour transcription rates keep falling while accuracy keeps climbing, which means the sticker price matters less every year and effective cost matters more. The teams that win the cost game aren't the ones chasing the lowest rate card — they're the ones who measure rework, consolidate their stacks, and buy accuracy where accuracy pays. A flat $4.50/hr for an entire voice pipeline would have sounded impossible two years ago. It's the baseline now, and the next surprise is already being built.
See the full, itemized rate card. Every model and add-on, priced per second, no minimums. View AssemblyAI pricing or explore Voice AI solutions.
Frequently asked questions about speech-to-text pricing
How much does speech-to-text cost per hour of audio in 2026?
Modern speech-to-text APIs run roughly $0.15 to $0.45 per hour, billed per second. AssemblyAI is $0.21/hr for async Universal-3.5 Pro, $0.45/hr base for streaming Universal-3.5 Pro Realtime, and $0.15/hr for the Universal-2 value tier.
What's the real difference between per-minute and per-hour billing, and why does it matter for my bill?
They're the same rate expressed differently, but per-minute providers often round each request up to a minimum, which inflates costs for short utterances. Per-second billing with no minimums, like AssemblyAI's, charges only for the audio you actually process.
How do accuracy rates affect the total cost of speech recognition, not just the API rate?
Accuracy sets your effective cost. A model at 85% accuracy needs 15–20 minutes of human correction per hour of audio versus about 5 minutes at 95% — a rework gap far larger than any per-hour rate difference between providers.
How much does it cost to build a voice agent, and can I consolidate STT, LLM, and TTS pricing?
Yes. Instead of paying three separate meters, AssemblyAI's Voice Agent API bundles speech-to-text, an LLM, and text-to-speech into one WebSocket at a flat $4.50/hr, billed per second — roughly 4x cheaper than OpenAI's Realtime API at about $18/hr.
How much extra does medical or clinical transcription cost on top of the base rate?
Medical Mode is a +$0.15/hr add-on (domain: "medical-v1") on Universal-3.5 Pro, so clinical audio runs about $0.36/hr all-in. It supports English, Spanish, German, and French and improves capture of drug names, dosages, and procedures.
When is it actually worth switching speech-to-text providers instead of staying put?
Switch when rework costs exceed your API bill, when you need capabilities your provider lacks (streaming, diarization, code-switching, native voice agents), when a three-vendor voice stack exceeds $4.50/hr, or when unpredictable pricing makes forecasting impossible.
Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
Related posts
