Speech-to-Text API Cost: Estimate Your Workload

November 3, 2026

A speech-to-text API budget starts with audio duration, but the useful number is the cost of delivering an acceptable transcript. Model rates, billable rounding, optional features, retries, human correction, and payment fees can all change that number.

This guide shows how to estimate those costs and how to read Speechall’s credit quote. All numeric rates in the worked examples are hypothetical planning inputs, not current provider prices or a Speechall offer. For current model information, use the model catalog; for your current top-up terms, sign in and review the billing quote.

Build a speech-to-text cost estimate

Start with a small worksheet. Record the source of each assumption and its date so that another engineer can reproduce the estimate.

The starting formula is billable minutes × price per minute. If a rate is quoted per hour, divide it by 60 before comparing it with a per-minute rate. Keep currencies separate until you apply a dated exchange-rate assumption.

Do not substitute recording length for billable duration without checking the model’s terms. For example, a two-channel recording might be treated differently from a mono recording. A streaming session may have different billing rules from a file transcription. These are questions to verify for the selected model, not properties shared by every speech API.

Worked example: monthly transcription volume

Imagine an application transcribing 2,000 recordings per month, with an average length of 12 minutes. Assume the selected model bills by the minute with no additional rounding in this example.

• Source audio: 2,000 × 12 = 24,000 minutes.
• Hypothetical transcription rate: $0.006 per minute.
• First-pass transcription cost: 24,000 × $0.006 = $144.
• Planning assumption: an additional 5% of source minutes are reprocessed and billed at the same rate.
• Additional processing: 1,200 × $0.006 = $7.20.
• Estimated transcription usage: $151.20.

This example assumes all additional attempts are billable. It does not assert that Speechall or a particular provider bills failed requests. Reconcile actual outcomes and usage records before replacing the assumption with a measured rate.

If storage and transfer add a hypothetical $12, and reviewers spend three hours at an assumed $25 per hour, total operating cost becomes $238.20, before payment fees and tax. Human review accounts for $75 of that estimate. A cheaper model that requires substantially more review could raise the total cost.

Compare quality at the same workload

Suppose a second hypothetical model costs $0.004 per minute. At 25,200 processed minutes it costs $100.80, saving $50.40 in model usage compared with the first example. If it requires three additional review hours at $25 per hour, the additional $75 of review exceeds that difference by $24.60.

This is a calculation, not a measured comparison between real models. To make the comparison useful, test the same recordings, required features, and acceptance criteria. Track correction time as well as transcript errors. Use the playground with a small, authorized audio sample before committing a large workload.

Account for short files and retries

Averages can hide billing effects. Under a hypothetical rule that rounds every recording up to a whole minute, 1,000 recordings of 10 seconds each represent about 166.67 source minutes but 1,000 billable minutes. At an assumed $0.006 per minute, that is $6 instead of approximately $1.00. Confirm the actual model’s billing increments before using either result.

Retries also need a deliberate policy. A client timeout does not prove that processing stopped. Blindly resubmitting a large file can create duplicate work. Record your application’s job identifier, model, attempt count, and outcome; investigate ambiguous results before retrying them. Check documented request behavior rather than assuming an idempotency header or automatic deduplication exists.

Budget separately for model evaluation. Sending the same hour of audio to four models means four hours of submitted audio, even though the original dataset is only one hour long.

How Speechall credit purchases work

Speechall’s current billing implementation accepts a base top-up amount in EUR, calculates a service fee in EUR, and converts the base amount into USD-equivalent usage credits. The billing screen states that one credit equals $0.01 USD equivalent. The service fee is separate from the principal used to calculate credits.

The service fee calculation uses the greater of:

• A percentage of the base amount, rounded to EUR cents, plus a fixed fee.
• A configured minimum service fee.

The minimum top-up, percentage, fixed fee, minimum fee, and exchange rate are supplied by the active billing configuration and quote. They are not fixed rates published by this article. The checkout subtotal is the base amount plus the service fee; applicable tax is calculated at checkout. Review the final checkout total before paying.

Worked example: a hypothetical top-up

Assume, solely to illustrate the calculation, that a valid base top-up is €100, the service fee is 5% plus €0.30, the minimum service fee is €1, and the quote uses 1 EUR = 1.10 USD.

These figures do not describe current Speechall fees or an available exchange rate. In the implementation, the converted principal is rounded down to whole USD cents. For a real purchase, the live billing quote shows the applicable minimum, fee breakdown, exchange rate, credit amount, and subtotal. Checkout calculates its own fresh quote, so review the final amount there too.

Do not compare a EUR checkout payment directly with a USD per-minute model rate. First identify the credits purchased, then estimate the usage those credits cover using the selected model’s current pricing. Prepaid cash expenditure and usage expense can occur in different months.

Set controls before scaling

Set a monthly usage target and a separate cash budget for credit purchases. Review usage regularly, compare processed minutes with your job records, and investigate unusual changes in cost per accepted transcript.

For an initial rollout, limit job concurrency and retries in your own application, test long recordings and short clips, and retain enough metadata to diagnose duplicates. Only collect the audio and transcript data your application needs. A cost investigation should not become a reason to log sensitive recordings indiscriminately.

If you are deciding how much integration work to own, read Speech API gateway vs direct integration. The getting started guide explains how to create your first request once the budget assumptions are clear.

Frequently asked questions

How much does it cost to transcribe one hour?

Multiply the current per-minute rate by 60, then account for the model’s billing increments, required options, and any separately priced services. At a hypothetical $0.006 per minute with no other charges, one hour costs $0.36. That example is not a current model quote.

Does a lower per-minute rate always reduce my costs?

No. Measure accuracy on your own audio, correction time, retries, and required features. Compare the cost of an accepted transcript, using the same workload for each candidate.

Are Speechall credits purchased in USD?

The current top-up flow takes a EUR base amount and converts that principal into USD-equivalent credits at the quoted exchange rate. The EUR service fee is added separately. Check the billing screen and final checkout for the current amounts and applicable tax.

Where can I see the current fee and minimum top-up?

Sign in to billing and enter a proposed base amount to view the quote. This article intentionally does not publish a fixed percentage or minimum that could become stale.

What is a useful first step before buying a large credit balance?

Create an API key, choose a model from the catalog, and run a small representative test in the playground. Review the resulting usage and transcript quality, then use the current billing quote to plan the next top-up. Transcription requests consume usage credits.