AI Call Transcription: Priced to Four Decimal Places, Accuracy Not Published

Research date: 23 September 2026. Last updated: 23 September 2026.

You can buy speech-to-text for $0.0043 a minute and know that figure to four decimal places. What you cannot find out from the same page is how often it gets the words right.

That is the whole problem with this category. Transcription is the layer everything else in an AI call centre stands on — if the words are wrong, the agent's decision is wrong, the CRM note is wrong, and the QA scorecard is scoring a conversation that did not happen. It is also the layer nobody publishes a number for.

And accuracy is not one number anyway. The best peer-reviewed measurement of commercial speech recognition found error rates almost twice as high for one group of speakers as another, on the same phrases, traced to the acoustic models themselves.

Brilo AI publishes this page and sells an AI voice agent, which has speech recognition inside it. Brilo does not publish a word error rate or name its speech-to-text provider either. That is stated again at the bottom, where it belongs.

How we compiled this

Documented research. Prices were read from vendors' own pricing pages on 23 September 2026. The accuracy research is a peer-reviewed study, cited with its limits. No transcription service was benchmarked by us, so this page contains no accuracy claim of our own — which is rather the point.

What transcription actually costs

Sources: deepgram.com/pricing, assemblyai.com/pricing, read 23 September 2026. Deepgram offers $200 in free credit, AssemblyAI $50.

Two things follow from this table, and the second matters more.

Transcription is cheap. At roughly half a cent a minute, the speech-to-text layer is a rounding error next to the rest of a voice deployment. AssemblyAI's own all-inclusive voice agent rate is $0.075 a minute, so transcription is somewhere around 5% of what the conversation costs you. If you are optimising your AI call centre spend, this is not the line to squeeze.

Nobody publishes accuracy. Both vendors above price their models to four decimal places and name several of them. Neither publishes a word error rate, an accuracy percentage, or any performance figure at all on the pricing page. You can compare cost precisely and quality not at all.

Accuracy is not one number

Word error rate is the standard measure: the proportion of words the system gets wrong. A single WER figure implies transcription has one accuracy, the way a car has one top speed. It does not.

The most rigorous published measurement of this is Koenecke et al., in the Proceedings of the National Academy of Sciences. They tested five commercial ASR systems — from Amazon, Apple, Google, IBM and Microsoft — on 19.8 hours of interviews with 42 white and 73 black speakers, matched on age and gender across five US cities. All five showed substantial disparities: an average WER of 0.35 for black speakers against 0.19 for white speakers. The authors traced this to the acoustic models rather than to vocabulary or subject matter, because the gap was equally large on identical phrases spoken by both groups. Source: Koenecke et al., "Racial disparities in automated speech recognition", PNAS, 2020. Read 23 September 2026.

The limits of that study, in the same breath as the figure. It is from 2020, and the five systems tested are not the models any of these vendors sell today; recognition has improved substantially since. It is also US English specifically. What has not changed is the structure of the problem: accuracy varies by who is speaking, the variation is a property of the model, and no vendor in this market publishes a breakdown that would let you check it for your callers. A single WER, if one were published, would still be an average hiding exactly this.

The operational translation is uncomfortable and simple. If your callers include a lot of people whose speech is under-represented in training data — regional accents, second-language English, older speakers, anyone with a speech difference — your real accuracy is not the benchmark accuracy, and the system will fail those callers more often than the others.

Phone calls are the hard case

Most published speech-recognition benchmarks use clean, wideband, studio-ish audio. A phone call is the opposite of that, in four specific ways.

It is narrowband. Traditional telephony carries roughly 300 Hz to 3.4 kHz, sampled at 8 kHz — a fraction of the frequency range of the microphone audio benchmarks are built from. Information the model could use is simply not on the line.

It is compressed. Codecs discard detail to save bandwidth, and what they discard is not chosen with transcription in mind.

It is noisy and uncontrolled. Callers ring from cars, streets, hospital corridors and speakerphones. Nobody is holding a microphone properly.

People interrupt, overlap and trail off. Benchmark audio is mostly one person speaking in complete sentences. Real calls are not.

A WER measured on clean audio therefore tells you very little about your phone line, and the gap runs in only one direction.

What to measure instead, and how

Since no vendor will hand you a number that applies to your callers, produce one. This is a day of work and it is the only way to know.

Take fifty of your own recordings, sampled across your actual call mix rather than the easy ones. Include your hardest accents, your noisiest lines and your most confusing product names.

Transcribe them and count the errors yourself. WER is countable by hand on a sample this size: substitutions plus deletions plus insertions, divided by the number of words actually spoken.

Stratify the result. Report it by accent, by region, by line quality — not as a single average. The average is the number that hides the problem.

Then measure entity accuracy separately, because it matters more than WER. A transcript that is 96% correct overall but gets the postcode, the order number or the caller's surname wrong has failed the call completely, while one that fumbles several filler words has not. Count how often the system gets names, numbers, addresses and reference codes exactly right. Almost nobody does this, and it is the measurement that predicts whether the deployment works.

Repeat it quarterly. Providers update models without telling you, and accuracy can move in either direction between one month and the next.

What to ask a vendor

Four questions, in writing, because none of this is on a pricing page.

1. Which speech-to-text provider and model do you use, and will you tell me when you change it?

2. What word error rate do you measure, on what audio, and can I see the breakdown by accent or speaker group?

3. Is the model tuned for 8 kHz telephony audio, or for wideband?

4. Can I run my own recordings through it before I sign?

An unwillingness to answer the first is the most informative result. Many voice products are assembled from third-party components, which is not a criticism — but it means the accuracy you experience is set by a vendor you are not in a relationship with, and that changes when they change it.

Brilo's position, and what it does not publish

Brilo publishes a free tier at 10 minutes a month, Pro at $149 for 600 minutes with overage at $0.16, Growth at $499 for 2,500 minutes at $0.14, and a custom tier above 5,000 minutes. Source: brilo.ai/pricing, read 23 September 2026.

What it does not publish: Brilo does not name its speech-to-text provider and does not publish a word error rate, an accuracy figure, or any breakdown by accent or speaker group. Nor does it publish retention or export terms for transcripts. On a page whose argument is that buyers should demand exactly those numbers, it would be dishonest not to say that Brilo is currently among the vendors who do not supply them.

The four questions above are the right ones to put to Brilo as well. Ask them, and ask to run your own recordings through it first.

Related reading

For scoring the conversations once you can read them, see AI call centre quality assurance — transcription is what makes QA possible at all. For where this sits in the per-minute cost, see AI call centre costs, pricing and ROI.