AI Voice directory for creators and small businesses

What speech actually costs, September 2026

ElevenLabs' pricing page now states text to speech burns 1 credit per character, which makes every vendor here comparable in one density-free unit: dollars per million characters. Ranked that way the field spans about 50x — $4.00 on Amazon Polly Standard, $15.00 on OpenAI tts-1, $30.00 on Deepgram Aura-2 and Murf's API, $50.00 on Cartesia Pro, and $200.00 on an ElevenLabs Starter subscription. The result that surprises buyers: ElevenLabs' own developer API at $50.00 per million is four times cheaper than its own entry plan, so the biggest available saving involves no vendor switch. Full ranking and sources: ElevenLabs alternatives, ranked by cost per character.

The door you buy through moves the bill more than the vendor

"AI voice" is four purchases on four meters — narration from a script, dubbing an existing recording, a live voice agent, and transcription — and no tool here wins all four. The decisive finding is that the same vendor charges very differently depending on how you buy: converted at Pro's $0.000165 credit value, ElevenLabs' subscription and API rate cards match to the cent on dubbing and music, run 1.65x higher by subscription for text to speech, and about 15x higher by subscription for transcription — $3.27 an hour of credits against $0.22 on Scribe v2. Job-by-job tables and picks: what a minute of AI voice actually costs.

Some vendors bill the script, others bill the audio

Every rate on this page assumes a meter, and the two most common ones measure opposite ends of the same job. ElevenLabs, Cartesia, Deepgram, Speechify and the Murf API bill characters going in, so the price is fixed the moment you paste the script. Murf Studio bills Voice Generation Time — the running length of the audio it produces — so the same script costs whatever it happens to run to. Priced both ways, a 1,000-word script is about $0.91 on ElevenLabs Creator annual and $0.95 or $1.45 on Murf Creator annual, because Murf's own help centre publishes two speaking paces 52% apart. The break-even lands between roughly 132 and 192 words a minute, which is inside the normal range of narration: ElevenLabs vs Murf works it through, including the rule that makes Murf cheaper from your second render onwards.

A voice clone is not sold by the hour — it is rationed

Nothing in this directory charges a line-item fee for cloning a voice. What you buy is a slot and a permission, and the permission has no price at all. ElevenLabs allocates zero professional clones on Free and Starter, one on Creator at $22 a month, one on Pro at $99, three on Scale at $299 and ten on Business at $990 — so above Creator a slot costs a flat $99 a month at every tier, and Creator is 77.8% cheaper per slot than the identical single slot on Pro. Cartesia runs the same ladder more cheaply at the top, with four clones on its $299 Scale tier against ElevenLabs' three. The rule that decides the shortlist before any of that, though, is whose voice it is: ElevenLabs states that "Even with their consent, you cannot clone someone else's voice" at its professional tier, and no plan lifts it. Slots, consent mechanisms and audio requirements: what AI voice cloning costs.

On one vendor here, paying more buys less

Every ladder on this page so far is a plan ladder. Amazon Polly has no plans at all — just four engines billed per character at $4.00, $16.00, $30.00 and $100.00 per million — and its ladder inverts in a way no pricing page shows, because coverage is documented somewhere else. The $4.00 Standard engine reaches 29 language variants in every Polly region; the $100.00 Long-form engine is six voices in two languages in a single region, us-east-1. In between, Generative costs 3.3x less than Long-form and beats it on voices, languages and regions, losing on exactly one thing: it cannot produce word timings at all. And on AWS's own published testing the cheap-ish $16.00 Neural engine has the lowest critical error rate of the three measured, at 0.77% of prompts against 1.79% for Generative. Full card, quotas and caveats: the Amazon Polly review.

One meter here has no per-minute price at all

Every rate on this page is a number you can multiply by minutes. OpenAI's Realtime API is the exception, and it is worth understanding before anyone budgets a voice agent from a headline figure. It bills audio by the token, and OpenAI's own cost guide states that the entire conversation is re-sent as input on every turn, so the meter compounds with call length. Modelled on a plain ten-minute call, the first minute costs about $0.048 and the tenth about $0.307 — 6.4x — and the whole call averages $0.178 a minute, roughly 3x Cartesia's managed $0.060. If prompt caching holds, the identical call averages $0.050 instead, a 3.58x swing decided by a mechanism OpenAI describes as best-effort. So the cheapest first minute in this category can be the dearest hour: the OpenAI audio API review shows the working.

The one job where price barely separates the vendors

Transcription is the commodity corner of this category. An hour of audio costs between $0.10 and $0.61 across every self-serve API CartSignal could price on 20 September 2026 — a 6.1x spread, against roughly 30x for narration and 12.8x for dubbing. So three other things decide the bill instead. Features are often charged separately, and AssemblyAI publishes more than a dozen per-hour add-ons that take a normal meeting transcript from $0.21 to $0.36. The streaming premium is not a constant: it is 0% at AssemblyAI, 11.6% at Deepgram, 77.3% at ElevenLabs and 278% at OpenAI. And a rounding rule can outweigh all of it — Google documents that "each request is rounded up to the nearest increment of 15 seconds", so a three-second utterance bills as fifteen, a 5x multiplier that no rate card shows. Full ranking and the add-on arithmetic: what AI transcription costs per hour.

One vendor here bills the clock, not the audio

Every per-hour rate in this category assumes you are buying audio. On AssemblyAI streaming, you are not. Its documentation states that "Streaming Speech-to-Text is billed on the total duration that your WebSocket connection stays open, not on the amount of audio you send", and that "You're charged for idle time on an open session just the same as you are for time when audio is actively flowing" — so the published $0.15 is a floor you only reach if somebody talks non-stop. A live session carrying speech half the time costs 2x the rate on the card; at 25% it is 4x, which is a bigger multiplier than the entire 6.1x price spread of the transcription market. Two further edges: concurrent sessions bill in parallel, so AssemblyAI's own example has a dual-streamed five-minute call billing as ten, and a session nobody terminates auto-closes after three hours and bills for all three. Its pre-recorded product is metered the opposite way and is unusually fair — billed to the exact second, with errored requests not charged at all. The same page also shows price and coverage running backwards: the $0.21 flagship reaches 18 languages, the $0.15 model beneath it reaches 99. The AssemblyAI review has the full card.

Frequently asked questions

Are these hands-on reviews?

No. These entries are public-data reviews unless specifically labeled as tested. Every rate, allowance and limit is transcribed from the vendor's own pricing page or documentation on the date recorded in each profile, and no audio or video is generated to assess quality. Where a vendor publishes its own benchmark figures, those are reproduced as vendor claims rather than independent measurements.

What is the first CartSignal vertical?

AI voice and AI video tools for creators and small businesses. The voice side covers narration, dubbing, live voice agents and transcription; the video side covers avatar video and generative video. Tools outside those two verticals are not profiled here.

Publishing AI voice in the EU

Since 2 August 2026, EU AI Act Article 50 requires generative audio to be marked machine-readably, and requires deployers to disclose synthetic audio that would falsely appear authentic. ElevenLabs began SynthID audio watermarking in June 2026; several other vendors here publish no clear statement. See labeling AI video and voice under the EU AI Act.

Methodology

CartSignal checks official product/pricing pages when reachable, public category signals, disclosure status, buyer-fit limits, and AI-readable structured data. Pages are labeled public-data review unless CartSignal has actually tested the product.