The Best Text to Speech APIs in 2026

Every text-to-speech API turns a string into audio. They do not bill the same unit, and that is the whole decision. OpenAI and Google Cloud charge per character. Cartesia and Hume charge in credits that convert to streaming minutes.

ElevenLabs charges credits too, but a character can cost half a credit or a full one depending on the model.

A comparison that lines a per-character rate up against a per-credit rate as if they were the same kind of dollar is how a team picks the wrong API for a million-user app.

Toolradar data: the directory homepage lists 10,796+ tools across 401+ categories, checked 25 September 2026.

Its free AI voice guide separately evaluated 15 AI voice tools to build that ranking, checked 25 September 2026, so open ElevenLabs, Deepgram, Cartesia, and Hume there for the wider AI voice category and use the ladders below for the API rate.

How we ranked: eight APIs, each read on the vendor's own pricing page on 25 September 2026.

The order follows the published dollar rate for a production workload, what a character or credit includes, and whether the fast path to a working voice agent needs a sales call or a card.

A voice cloning app or a dubbing tool is a different buy than a raw API.

That comparison lives in the free text-to-speech tools guide on Toolradar.

A text-to-speech API takes a string, a voice ID, and some settings, and returns audio, usually in a few hundred milliseconds for a short clip.

The bill almost always follows the input text, not the audio length, which is why a script full of SSML tags and punctuation costs more than the same words written plainly.

The category splits by billing shape.

Per-character APIs (OpenAI, Google Cloud, Azure, Amazon Polly) post a flat rate per 1M characters and are the easiest to forecast.

Credit-based APIs (ElevenLabs, Cartesia, Hume) sell a monthly bundle where a character or a minute consumes a variable number of credits depending on the voice model, so the effective per-character cost moves with the model you pick.

Deepgram sits in between: a flat rate per 1,000 characters with a discounted annual commitment. A voice agent built on any of these usually sits behind an API gateway and a backend that logs every call for billing.

A 10 million character month, roughly a long audiobook catalog or a year of a busy voice agent, is a different invoice on every API. On Google Cloud's Standard voices it is $40, most of it inside the 4 million free characters.

On OpenAI's tts-1 it is $150. On Azure's custom Professional Voice it is $240.

On ElevenLabs' pay-as-you-go v3 rate it is close to $1,000, the price of the voice most consumer apps actually ship because it is the one that does not sound synthetic.

Devshot data: 28,000 developer readers get this brief, and the audience file updated 20 September 2026 puts the open rate at 30%.

They forecast an API the way they forecast a cloud bill, so a credit system that quietly halves or doubles the per-character cost by model is the month-two surprise. Later pricing changes land in the Devshot brief for the next check.

Latency is the other axis.

A chatbot reading a paragraph can tolerate a full round trip. A phone agent or a live game NPC cannot, and that is a separate buy from a batch narration job.

The monitoring tools guide, the Datadog review, and the error tracking tools guide cover the dashboards that catch a voice API's slow p99 or a failed synthesis call before a user does.

Compare the flat-rate APIs first, because they convert cleanly to a per-1M-character number. Google Cloud's Standard and WaveNet voices are the cheapest ongoing free tier here.

OpenAI's tts-1 and Azure's Neural voices land on the same flat rate, and tts-1-hd doubles it for higher fidelity.

Amazon Polly's Standard tier matches Google Cloud's floor, but its Neural and Long-Form tiers climb well past it, the second-most expensive rate on this page after Google's Studio voices.

Deepgram's Aura-2 pay-as-you-go rate drops further on a Growth plan that needs a $4,000-plus annual prepay.

The credit-based APIs do not convert as cleanly. ElevenLabs' API rate works out to roughly ten times OpenAI's tts-1 rate on the v3 model, or half that on v3 Conversational, while its monthly plans run $6 to $990.

Cartesia and Hume bill in minutes and characters respectively inside a monthly credit bundle: Cartesia from $5/mo for about 133 TTS minutes, Hume from $3/mo for 30,000 characters, with Hume's Creator tier at $14/mo regular ($7 the first month).

Azure's custom Professional Voice adds a separate synthesis rate plus $52 per compute hour to train the model, verified on each vendor's pricing page, September 2026.