Research-backed calculator guide

AI Voiceover for Video: Budget Narration, Mixing and Review per Finished Minute

Estimate AI voiceover cost per finished minute by separating character-metered synthesis from rerenders, pronunciation fixes, mixing, review, and usage assumptions.

What “finished minute” means

A finished minute is one minute of the audio that you actually deliver after trimming silence, correcting pronunciation, choosing takes, applying basic cleanup, balancing levels, and reviewing it against the video. It is not one minute of an API request, one minute of editing time, or one minute of a subscription allowance.

This distinction matters because two projects with the same final runtime can consume different inputs. A calm explainer may need one generation and a quick check. A medical script with names, abbreviations, and several retakes may use far more characters and review time. A video with music, sound effects, multiple speakers, or tight sync has additional layers that should not be hidden inside a “voiceover per minute” headline.

How to calculate AI voiceover cost per finished minute

Use this sequence for each project:

1. Measure the script. Count words and, preferably, the actual characters sent to the text-to-speech endpoint. Include the punctuation or markup that your provider bills.
2. Convert the script to an initial audio estimate. For planning only, 130–170 spoken words per minute is a useful range for English narration. Use a timed read when pacing matters.
3. Add production reserve. Multiply characters by a rerender factor for pronunciation fixes, bad emphasis, timing changes, and small script edits. A 15% reserve is an illustrative assumption, not a market standard.
4. Apply the vendor meter. Character-priced products should be calculated as billable characters divided by 1,000 or 1,000,000, multiplied by the published price. Do not compare a per-character rate directly with a bundled conversation-minute rate.
5. Add workflow labor. Include script QA, voice selection, pronunciation markup, generation monitoring, cleanup, mix, sync check, and client review. If you do the work yourself, use an internal hourly cost rather than pretending it is free.
6. Add project-specific extras. These can include music, sound effects, translation, dubbing, extra speakers, rush work, usage rights, storage, taxes, and video editing.
7. Divide by final runtime. Total project cost divided by the final edited minutes gives the cost per finished minute.

The formula is:

cost per finished minute = (TTS usage + workflow labor + extras) / final edited minutes

For TTS usage:

TTS usage = (input characters × rerender factor / billing unit) × vendor rate

When a provider uses monthly credits, allocate only the portion of the subscription consumed by this project. If a $22 plan supplies 275,000 characters and your project consumes 6,038 characters, a simple usage allocation is $22 × 6,038 / 275,000 = $0.48. That is an accounting allocation, not necessarily the amount charged at checkout, because the subscription may be purchased for other projects and unused capacity may expire or roll over under its own terms.

What current vendor meters tell you

As checked on November 2, 2026, the official ElevenLabs API pricing page lists text-to-speech models at $0.08 per 1,000 characters for v4 and $0.04 per 1,000 characters for v4 Turbo. The page also shows model-specific included-character allowances and warns that the approximate minute conversion is based on typical usage. Use the character meter for the primary calculation, and treat the displayed minute figure as a planning hint. ^[1]

Deepgram's official pricing page states that text-to-speech is billed per 1,000 input characters. It lists Aura-2 at $0.030 per 1,000 characters on pay-as-you-go and Aura-1 at $0.0150 per 1,000 characters, with different Growth rates. The same page separately lists speech-to-text, voice-agent, add-on, and concurrency meters. A voiceover estimate should include only the layers you actually use. ^[2]

AssemblyAI's official pricing page lists its Text-to-Speech API at $15 per 1 million input characters, with cloned voices and EU processing at the same price. That is a vendor quote for the TTS API layer, not a promise that the final video will cost $15 per million characters: review, retries, editing, and other services remain separate assumptions. ^[3]

Those figures are not a universal ranking. They use different models, voice qualities, plans, limits, and commercial terms. They also do not include your time. If you are comparing several models, the Model pricing comparison page is a useful place to organize the provider meter, unit, plan, and scenario rather than comparing the smallest headline number.

Worked example: a five-minute finished video

Assume a US English explainer with one AI narrator and a final runtime of five minutes. The script has 750 words, or 150 words per finished minute. For a planning estimate, assume seven billable characters per word including spaces and punctuation. That gives 5,250 initial characters. Add a 15% reserve for rerenders and pronunciation fixes, producing 6,037.5 billable characters, rounded to 6,038.

Assume the operator spends four minutes on script preparation and voice selection, three minutes on pronunciation checks and regeneration, and five minutes on cleanup, mix, and final review. That is 12 minutes of labor. At an internal cost assumption of $40 per hour, labor is $8.00 for the project, or $1.60 per finished minute. The $40 rate is an explicit planning assumption, not a vendor quote.

The striking result is not that one provider is always cheapest. It is that the synthesis line is small relative to a modest amount of human workflow. If the script is clean and the process is automated, labor may fall sharply. If the video needs detailed pronunciation work, multiple voices, or frame-accurate changes, labor can dominate even when the TTS invoice is tiny.

For a one-minute project, do not simply divide the five-minute labor estimate by five and assume the same economics. The same fixed setup can make a short deliverable more expensive per minute. Conversely, a 30-minute course may amortize voice selection and project setup across many minutes, while increasing the chance of script corrections. Model both a short-job case and a steady-volume case.

Subscription allocation versus pay-as-you-go

A monthly plan is a capacity purchase, not automatically a per-finished-minute price. Suppose a plan costs $22 and includes 275,000 characters. At seven characters per word and 150 words per minute, its rough theoretical capacity is about 262 finished minutes before rerenders. A 15% reserve lowers usable planning capacity to about 228 minutes. The subscription allocation is therefore approximately $22 divided by 228, or $0.10 per planned finished minute, before labor.

That allocation is useful only if the plan is actually consumed by this work. If you buy the plan for several clients, allocate characters by project. If the plan includes commercial rights, extra voices, or other products, decide whether those benefits belong in the project cost. If unused credits expire, include expected utilization rather than the maximum headline allowance. A pay-as-you-go plan can be more economical for sporadic work, while a subscription can reduce effective cost at predictable volume.

Do not confuse the voiceover calculation with token billing for a script-writing model. If an LLM drafts, rewrites, or translates the narration, its input and output tokens are another line item. The AI token cost calculator can be used for that separate estimate. If you are combining multiple providers or model families, the LLM cost calculator helps keep token meters separate from audio characters.

What to include, exclude, and verify

Include the text sent to TTS, markup that the provider counts, failed or replaced generations, pronunciation dictionaries, review time, cleanup, loudness balancing, file exports, and the share of any required subscription. Include a contingency when the script is not final. A small correction after approval can trigger a new generation, a new mix, and a new review pass.

Exclude items only when you state the exclusion. Common exclusions are music and sound effects, video editing, captions, translation, dubbing, live direction, human pickup talent, custom voice training, paid advertising usage, legal review, storage, tax, and platform fees. A “voice-only” estimate should not be presented as a complete video-production quote.

Verify the commercial license for the exact plan and voice. A free tier may be suitable for auditions but not for a monetized client deliverable. Also verify whether a cloned or custom voice has separate permissions, whether the voice may be used in advertising, and whether your contract covers the intended geography and duration. The vendor price pages are authoritative for their listed meter and rate at the check date, but your terms may differ by account, region, volume, or negotiated agreement.

FAQ

Is AI voiceover cost usually quoted per finished minute?

Not consistently. Human narration is often quoted by finished minute, word band, session, or usage rights. AI platforms more commonly meter characters, credits, subscription allowances, or conversation minutes. You can create a finished-minute estimate by calculating the full project total and dividing by the final edited runtime, but you should preserve the underlying meter so the estimate can be audited.

How many characters are in one finished minute of narration?

There is no fixed number. A planning range of about 130–170 spoken words per minute is more defensible than assuming every voice reads at the same speed. Character count then depends on language, word length, punctuation, spaces, numbers, SSML, and pauses. In the worked example, 150 words and seven characters per word produce 1,050 characters per finished minute before the rerender reserve. Replace that assumption with a count from your real script before purchasing capacity.

Does the AI TTS API price include editing and mixing?

Usually, an API price covers the metered synthesis request, not your full delivery workflow. Even when a studio product includes editing features, confirm what is included in the plan and what is merely available in the interface. Budget separately for pronunciation fixes, selecting takes, removing unwanted silence, cleanup, level matching, music ducking, sync checks, exports, and client review unless your quote explicitly bundles them.

Is a subscription cheaper than pay-as-you-go for a single video?

Only if the project uses enough of the included allowance and the plan's license fits the use. Allocate the consumed characters and include a realistic rerender reserve. For a single five-minute video, a low-cost pay-as-you-go meter may be easier to explain. For recurring weekly production, a subscription may lower the allocated TTS cost, but unused credits, commercial terms, plan limits, and labor can matter more than the raw character rate.

What makes AI voiceover cost rise fastest?

The largest practical drivers are script changes after generation, difficult pronunciations, multiple speakers, translations, tight video sync, rush review, and usage or licensing requirements. A verbose script raises character usage, but it may still be a small share of total cost. Repeatedly regenerating whole chapters for one changed line is especially wasteful; isolate and replace the affected segment when the workflow allows it.

Should I compare vendors on price per minute?

Use price per character or the vendor's actual billing unit first, then convert to a project-specific finished-minute estimate. Vendors can use different assumptions for speaking speed, included characters, markup, pauses, and plan capacity. A bundled voice-agent minute may include several layers, while a TTS character price covers only synthesis. Comparing unlike denominators can produce a false winner.