BenPicks
Search BenPicks

Tutorial · Evidence checked 2026-08-29

How to Calculate AI Voice Generation Cost

A practical, evidence-led decision guide. Product capabilities and limits are separated from anything that would require hands-on testing.

An AI voice quote is easy to read and surprisingly easy to misuse. One service bills characters, another finished minutes, another credits, and an editor may charge a flat subscription while premium voices consume the same allowance at different multipliers. The honest comparison starts by refusing to convert those units until a real job connects them.

Reproducible evaluation workflow with inputs, checks and decision record

Start with the deliverable, not the vendor unit

Write down monthly new words or characters, accepted finished minutes, languages, number of variants, expected correction rate and peak deadline. Count the people who generate, review and approve. Note whether the output is public, paid, broadcast or embedded in a product, because the commercial-rights tier can change the eligible plan.

The bridge between characters and minutes must come from your material. Generate a representative script and record both its submitted characters and accepted duration. Do not use a universal “characters per minute” constant for every language, voice and pacing style.

Keep six meters separate

  1. Input meter: characters, tokens or words submitted.
  2. Output meter: seconds or minutes generated, downloaded or accepted.
  3. Plan meter: credits, included allowance or projects.
  4. Quality meter: regeneration caused by defects or changes.
  5. People meter: review, pronunciation repair, editing and mastering time.
  6. Rights meter: the plan or licence required for the intended commercial use.

NaturalReader Commercial illustrates why the distinction matters. Its documentation applies different credit multipliers to voice providers: a voice that consumes allowance at 20× does not automatically cost 20× more than buying another service directly. It means the NaturalReader allowance disappears 20× faster for that selection. To claim a price multiple, you would also need the plan price, allowance, comparable direct-vendor rate and identical output volume.

A spreadsheet model that does not lie

Create one row per provider/model/voice combination, not one row per brand.

InputRecord
Monthly new characterssource scripts only
Regeneration rateobserved failed/changed characters ÷ new characters
Billable characters or minutesprovider meter after rework
Plan allowance and multiplierexact selected tier/voice
Overage or pay-as-you-gocurrent price and currency
Reviewer and editor hoursobserved time × loaded hourly cost
Add-ons / seats / storageonly required items
Rights-qualified?yes / no / unresolved

Calculate:

`billable input = new input × (1 + regeneration rate)`

`monthly workflow cost = subscription + overage + required add-ons + labour + hosting/integration`

`cost per accepted minute = monthly workflow cost ÷ accepted finished minutes`

Keep taxes and currency conversion outside the provider comparison unless the same purchase date and rate are used.

Worked example: character pricing with rework

Suppose a team submits one million characters and observes 25% regeneration. Billable input is 1.25 million characters. At Amazon Polly's published rates, synthesis alone would be $5 Standard, $20 Neural, $37.50 Generative or $125 Long-Form before free allowance. The engine choice changes this example more than the rework percentage.

Now add six hours of review at an internal loaded cost of $40 per hour. Labour contributes $240—more than any Polly engine charge in the example. That does not make synthesis price irrelevant at scale; it shows why a buying page that compares only API rates can point to the wrong saving.

Worked example: credits with a voice multiplier

Assume a plan includes 100,000 base credits. A 1× voice can consume the full 100,000-unit workload; a 10× selection exhausts the same pool after 10,000 comparable units; a 20× selection after 5,000. Those figures describe capacity, not the direct vendor's price.

To compare an aggregator with buying from the voice provider, calculate both routes symmetrically. The aggregator may add a unified editor, several providers, one invoice and a simpler approval workflow. The direct route may offer deeper controls and a better native rate. Put the saved operator time on the aggregator side and any integration work on the direct side.

Finished-minute and localisation workflows

Narakeet, Maestra, Respeecher and other production tools can meter different things: generated minutes, downloadable minutes, translation, dubbing, cloning or marketplace/API usage. Do not add them as if they were the same minute. A translated minute with lip-sync is a different service from one minute of plain TTS.

For multilingual work, split the model into transcription/source preparation, translation, TTS/dubbing, lip-sync, subtitles, review and regeneration. Multiply source duration by target languages only for stages that genuinely repeat. If lip-sync applies to three of ten languages, do not multiply it by ten.

Measure the correction rate instead of guessing it forever

Use at least three real projects. For each, record submitted units, accepted units, defect-driven regeneration and client/editorial changes separately. A vendor should not be blamed for a script change; the capacity model still must pay for it.

After the first month, replace the planning assumption with the median and a peak case from observed work. Keep the peak scenario if launches or seasonal production create bursts. A plan that covers the average but blocks the delivery week is under-sized.

Rights can make the cheapest plan ineligible

Before comparing totals, confirm commercial output rights, voice/persona consent, redistribution, client work, API embedding and any restrictions on synthetic impersonation. Preserve the exact terms or licence version reviewed. A $10 plan that does not cover the intended distribution is not the cheapest option; it is not an option.

The decision table to take into procurement

QuestionCandidate ACandidate B
Exact plan/model/voice
Baseline billable units
Peak billable units
Observed regeneration
Accepted finished minutes
Provider charge
Labour and integration
Cost per accepted minute
Commercial rightsconfirmed / openconfirmed / open

Do not collapse an unresolved rate, multiplier or licence into zero. Mark the total incomplete and ask the vendor or run the account-level check. A transparent blank is more useful than a precise-looking estimate built on the wrong unit.

Put two genuinely different routes into the sheet

For a character-metered API comparison, start with the exact engine prices in Amazon Polly and the selected-model rate in Google Cloud Text-to-Speech. For a commercial editor or aggregator decision, compare NaturalReader Commercial with the direct provider route and give the unified workflow, invoice and operator time the same attention as the credit multiplier. For finished-minute production, use Narakeet, Maestra or Respeecher only after defining which minute each product actually meters.

Enter the real monthly workload, preserve every unknown as an unknown and calculate a baseline plus a peak month. If one route remains incomplete, request the missing allowance or licence term before buying; if both totals are complete, trial the cheaper accepted-output workflow rather than the cheapest advertised plan.

Official sources used for the example boundaries