What an audio hour costs
All five options below are transcription APIs: you pay per minute of audio, with no monthly subscription. Prices come from the pricing pages of OpenAI, Google Cloud Speech-to-Text, Deepgram and AssemblyAI on 28 October 2026. For Deepgram and AssemblyAI these are the rates for pre-recorded audio, not live transcription. Hryvnia amounts use the National Bank of Ukraine rate for 28 October: UAH 45.6008 per dollar.
| Service and model | Per minute | Per hour | 10 hours a month | 100 hours a month |
|---|---|---|---|---|
| AssemblyAI Universal-2 | $0.0025 | UAH 6.84 | UAH 68 | UAH 684 |
| AssemblyAI Universal-2 with speaker labels | $0.0028 | UAH 7.75 | UAH 78 | UAH 775 |
| OpenAI gpt-4o-mini-transcribe | $0.003 | UAH 8.21 | UAH 82 | UAH 821 |
| Google Speech-to-Text V2, dynamic batch | $0.003 | UAH 8.21 | UAH 82 | UAH 821 |
| Deepgram Nova-3, single language | $0.0043 | UAH 11.77 | UAH 118 | UAH 1,177 |
| OpenAI gpt-transcribe | $0.0045 | UAH 12.31 | UAH 123 | UAH 1,231 |
| OpenAI gpt-4o-transcribe-diarize | $0.006 | UAH 16.42 | UAH 164 | UAH 1,642 |
| Google Speech-to-Text V2, standard | $0.016 | UAH 43.78 | UAH 438 | UAH 4,378 |
| Whisper on your own computer | — | UAH 0 in API fees | UAH 0 | UAH 0 |
AssemblyAI quotes an hourly price: $0.15 for Universal-2 plus $0.02 for speaker labels; the table converts this to a per-minute figure and rounds it. At Deepgram, speaker labels for pre-recorded audio are included in the price. Google's dynamic batch is five times cheaper because the result does not come back immediately.
6,000 minutes × price per minute × UAH 45.6008/$ (NBU rate for 28.10.2026). Rates checked on 28 October 2026 on the providers' pages. Excludes volume discounts and file storage costs.
Even the most expensive row is UAH 4,378 for 100 hours — that is 20 working days with five hours of recordings each day. For a small team, transcription is rarely the main cost. What matters more is whether the service recognises your language correctly and whether you may hand it these recordings at all.
Free allowances to start with: AssemblyAI gives up to 185 hours of pre-recorded transcription, Deepgram gives $200 of credit, and Google's V1 API does not charge for the first 60 minutes a month.
What about Ukrainian
Each provider describes Ukrainian support differently, and you only see it in the documentation, not on the pricing page.
Sources: OpenAI's speech-to-text guide, the Google Speech-to-Text V2 language table, Deepgram's models overview, AssemblyAI's supported languages page, the openai/whisper repository. We did not measure accuracy ourselves: the ratings are the providers' own figures.
Details the table does not show.
OpenAI recommends gpt-transcribe for new projects. There is no separate language list for it: the guide says the model accepts ISO 639-1 codes and returns the languages it could reliably detect. The code for Ukrainian is uk. Speaker labels exist only in gpt-4o-transcribe-diarize, and that model, together with whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe, is deprecated: they shut down on 26 February 2027.
Google supports uk-UA in the chirp, chirp_2, chirp_3, long and short models, but not in every region. For chirp_3 the language table marks automatic punctuation, and does not mark speaker labels for Ukrainian.
AssemblyAI released its newest model, Universal-3.5 Pro, for 18 languages, and Ukrainian is not one of them. Ukrainian is in Universal-2, and the company puts it in the "high accuracy" group — up to 10% word error rate by its own measurement. If you set the language manually and a feature is not supported for it, the API returns an error rather than silently skipping the feature.
Whisper is an open OpenAI model under the MIT licence. The repository lists about 6 GB of video memory for the fast turbo model and about 10 GB for large. Speaker labels are not among the capabilities in its README.
Four pricing traps
- A stereo call recording costs double at Google. Google bills each audio channel separately. A call recorded in two channels (agent and customer) counts as two minutes for every minute of conversation: 100 hours of standard processing become UAH 8,755 instead of 4,378.
- An hour-long meeting does not fit in one OpenAI request. The file limit is 25 MB. An hour of mp3 at 64 kbit/s weighs 28.8 MB, and 25 MB holds only 52 minutes. Speech can be compressed to 32 kbit/s mono — an hour then weighs 14.4 MB — or the recording can be split.
- Deprecated OpenAI models are still on the price list. Four models in the table above work until 26 February 2027. If you are building a process now, use gpt-transcribe so you do not have to redo the integration in four months.
- Deepgram's discounts apply to live transcription. The pricing page marks the promotional rates for streaming; the table above uses the regular rates for pre-recorded audio.
Meeting, voice notes or calls
The first question is not "which service is best" but what exactly you are transcribing. That decides whether you need speaker labels and whether the recording can go to the cloud at all.
- Could the recording contain health data, passport or card details, or trade secrets?A doctor's or lawyer's consultation, a deal discussion. A cloud service here means sending such data abroad. Yes → Whisper on your own computer
- Are several people speaking, and does it matter who said what?A meeting where tasks are assigned, or an interview. Without speaker labels, the minutes have to be marked up by hand. Yes → Deepgram Nova-3 or AssemblyAI Universal-2 with speaker labels
- Are these telephony calls recorded in two channels?The channels already separate agent and customer, so speaker labels are not needed, but a two-channel recording costs double at Google. Yes → check whether the service bills channels separately; the sales team calculation is in the article on call analysis
- Are these short voice messages from customers, one speaker each?The cheapest option that supports Ukrainian will do. Here the cost depends on the number of messages, not the model. Yes → AssemblyAI Universal-2 or OpenAI gpt-transcribe
The calculation for a sales team — how many call minutes, which metrics, and what the personal data law requires when recording a conversation — is in the article on AI in sales and call analysis.
Where recordings may go
A recording of someone's voice is their personal data. The general rule is the same as for any AI service: you need a legal basis for processing, a contract with the service, and an understanding of where the data is processed. A traffic-light guide to what is allowed is in the article on what data to give an AI assistant.
For the OpenAI API, the data guide says that since 1 March 2023 data sent to the API is not used to train models unless you opt in. Abuse-monitoring logs, however, are kept for up to 30 days by default. Zero Data Retention is enabled only after approval by OpenAI. Deepgram, AssemblyAI and Google have their own terms — read the data processing agreement before connecting.
Whisper on your own computer removes the question: the recording goes nowhere. The price is a graphics card, setup work and no speaker labels.
What AI costs for a whole team once you add assistant subscriptions to transcription is covered in the article on an AI budget for ten people.
Our position
For Ukrainian meetings, choose a service whose documentation lists Ukrainian together with speaker labels: as of 28 October 2026 that means Deepgram Nova-3 and AssemblyAI Universal-2. The price difference between them and the cheapest option is a few hundred hryvnias a month even at 100 hours, while marking up "who said what" in minutes by hand costs more. Before connecting, run 3–5 of your own recordings through two services and compare the texts: the accuracy figures on providers' sites were not measured on your recordings.
This position does not hold if the recordings contain customers' health data, passport data or trade secrets. A cloud service without a separate data processing agreement is then unsuitable at any price, and the only option is local Whisper, even with speaker labels added by hand.
Checked on 28 October 2026 on the pricing pages of OpenAI, Google Cloud Speech-to-Text, Deepgram and AssemblyAI, in OpenAI's speech-to-text and API data guides, the Google Speech-to-Text V2 language table, Deepgram's models overview, AssemblyAI's supported languages page, the openai/whisper repository, and against the NBU exchange rate on bank.gov.ua.