Text-to-speech APIs you can pay per call

What a text-to-speech API does, and who calls it

A text to speech API turns supplied text into audio. The caller chooses a script and, where supported, a voice, language, model, pace or expressive setting. Useful output can be an audio stream, inline file or asynchronous job that eventually provides a downloadable asset. Those delivery forms have different client requirements.

An agent can use synthesis to read a short reply, announce an alert or produce a reviewed telephone greeting. A narration pipeline needs to track the script revision and settings over longer passages. Before comparing prices, establish maximum input length, audio format and whether the application needs immediate playback or can wait for a background job.

The table covers 8 services with matching synthesis routes as of October 26, 2026. It excludes transcription, credit purchases, chat and topic-driven podcast production from the speech task. It also leaves out 0 candidates that pass the listing checks but lack a matching priced endpoint in the recorded list. A reference quote can still depend on the input: the character-priced and model-selected products need request-specific inspection before a script budget is approved.

What the 8 text-to-speech services actually return

The classification follows documented endpoint products. It counts services per type, so a service offering more than one product can appear in more than one row. Each quote comes from the cheapest priced endpoint of that type, with its source route attached.

Each price is the cheapest endpoint of that service for that kind of data: the same number wherever this page quotes it.

Use this table to narrow the product before comparing hosts. A matching type identifies an advertised output, not verified accuracy or complete coverage of your input. The request schema, source period and returned missing-value states still need to fit the job. Read the service-specific notes below for limits that a broad product label cannot express.

The 8 text-to-speech APIs and their matching endpoint prices

One row per service, ordered by measured buyers. The price names the selected task route rather than an unrelated utility on the same host. Primary types determine the preferred quote; a service without a primary type uses its cheapest matching task endpoint.

$/CALL is the price of the endpoint named under it: the cheapest text-to-speech audio endpoint of the service or, for a service without one, its cheapest speech-synthesis endpoint. Never the cheapest tool of a gateway. Other kinds of data have their own price in the tables below. SPEECH-SYNTHESIS ENDPOINTS counts the speech-synthesis endpoints out of all the endpoints the service lists. BUYERS 30D counts distinct wallets that paid the whole service on-chain in 30 days, not only its speech-synthesis endpoints. "Generalist" marks a gateway that sells more than three unrelated kinds of tools: it is never called the cheapest.

The buyer figure covers the host's measured payments, so it does not establish demand for this particular endpoint. A gateway can sell unrelated tools and a reseller can share an upstream source with another listing. Use these signals to investigate operational fit rather than inferring independent data collection or product quality from wallet counts.

How to choose a text-to-speech API

Input limits and cost units

Check whether a quoted request covers a short character band, a longer band or a calculation based on the actual text. Count punctuation and whitespace according to the provider's schema. A reference minimum is useful for discovering the endpoint but cannot establish the total cost of a long narration.

Audio format and playback

WAV and MP3 have different storage and playback implications. Inspect the sample rate, channel count, codec and response content type required by your player or telephone system. An inline payload needs decoding, a downloadable asset needs an accessible URL, and an asynchronous job needs state management before playback.

Voice and language fit

A declared language count or model name is a provider statement. Evaluate the accent, proper names, abbreviations, punctuation and number reading your application needs. Record a representative approved script and voice setting so future comparisons use the same material instead of unrelated demo clips.

Controls and delivery time

A speed setting, quality control and persona describe different knobs. Test their effect at the desired playback conditions. Monitoring latency describes our unpaid endpoint checks; it should not be labeled full synthesis time or time to first audio. For interactive speech, measure completion with an actual permitted test workflow before setting a user-facing timing promise.

Which endpoint fits which job

Start with the job, then compare the candidates whose classified endpoints match it. The price in a job table is tied to those types, which can differ from the service's selected price in the main table. Do not assume that one host's entry quote covers another product it sells.

Speak a short agent answer

Define acceptable time to first audio, clip format and maximum text length. Test punctuation, numbers and a proper name representative of your application. Keep the model and voice explicit so that a later provider default change does not silently alter the voice of a saved workflow.

Candidates today, most buyers first: Cluster Protocol ($0.015/call, generalist, 719 buyers) · GlianaAI ($0.014/call, generalist, 27 buyers) · ForgeMesh Voice ($0.001/call, 14 buyers) · kino402 ($0.007/call, 13 buyers) · Synap by Linkrra ($0.01/call, 7 buyers) · x402.eleeth.com ($0.25/call, generalist, 1 buyer) · GPUOps AI Inference ($0.01/call) · Voix Off TTS FR — French neural voiceovers pay-per-call ($0.03/call)

Read a longer paragraph

Select a documented longer-input tier and inspect character counting before splitting text. Preserve sentence boundaries and consistent settings across chunks. A long route is useful evidence of an input limit; it does not establish streaming, seamless concatenation or the duration of the produced clip.

Candidates today, most buyers first: ForgeMesh Voice ($0.003/call, 14 buyers)

Control pace or expressive style

Distinguish speaking speed from a named persona and from a model-quality setting. Record the chosen parameters with the script. Evaluate a sample at the intended playback speed, because faster delivery can reduce intelligibility even if the generated file meets the requested technical format.

Candidates today, most buyers first: ForgeMesh Voice ($0.003/call, 14 buyers)

Run asynchronous voice jobs

Persist the job identifier before checking status and model completion separately from request acceptance. Inspect documented expiry and failure states. A successful submission does not prove that the final audio is ready; the client needs a recoverable path from paid submission to the completed asset.

Candidates today, most buyers first: x402.eleeth.com ($0.25/call, generalist, 1 buyer)

Prepare a French telephone message

Check the telephone system's required sample rate, channel count and codec as well as the language. Listen to names, opening hours and phone digits. Confirm the short-message limit in the endpoint and keep an approved script revision so the deployed greeting corresponds to the intended wording.

Candidates today, most buyers first: Voix Off TTS FR — French neural voiceovers pay-per-call ($0.03/call)

Service details and limits to check

The notes below are read from the provider's description and endpoint list captured on October 21, 2026. They attribute advertised capabilities and source information; x402-list has not independently tested the accuracy of the delivered data or generated output. Price, uptime and buyer tokens come from the current page model.

ForgeMesh Voice

ForgeMesh Voice describes WAV synthesis in a short base tier, a longer base tier, controlled tiers and a persona tier. Speech starts at $0.001. Its long-text tiers start at $0.003 and describe 501 to 2,000 characters; the short tier describes up to 500 characters. Speed and quality controls start at $0.003, while the named-persona route costs $0.005. The descriptions advertise 31 languages and 20 voices: 10 standard voices plus 10 personas, with the persona options available through the Custom tier. Treat those as declared options rather than an evaluation of pronunciation. Select the tier by input length and control requirements before comparing its quote with another endpoint.

5 speech-synthesis endpoints of 5 · $0.001/call (GET /v1/tts/base) · Base · uptime 30d 100% · 201 ms · 14 buyers in 30 days

Synap by Linkrra

Synap by Linkrra lists /api/billing/x402/tts for WAV synthesis. Its recorded speech quote is $0.01, but the endpoint describes a character-dependent calculation with a minimum. That reference must not be treated as the final cost of an arbitrary script. Submit or quote the actual input according to the provider's schema and compare the returned payment requirement with the budget. Its chat and credit top-up routes are separate products and do not determine the speech price in this table. For a spoken agent reply, keep the source text and request settings with the audio so a mistaken name or abbreviation can be traced back to the input.

1 speech-synthesis endpoint of 3 · $0.01/call (GET /api/billing/x402/tts) · Base · uptime 30d 100% · 230 ms · 7 buyers in 30 days

x402.eleeth.com

x402.eleeth.com describes /api/tts as MP3 synthesis returned through an asynchronous job. The recorded speech quote is $0.25, and the job-oriented data type is priced at $0.25. Build the client around submission, status and completion rather than assuming the first successful request already contains playable audio. The same host also describes a podcast product, but topic-driven podcast creation is outside plain text synthesis here. Confirm retention, the final audio URL and failure states before using a job in an interactive experience. An average monitoring latency is a measure of the endpoint check, rather than the time needed to synthesize and download a complete script.

1 speech-synthesis endpoint of 34 · $0.25/call (POST /api/tts) · Base · uptime 30d 100% · 153 ms · 1 buyer in 30 days

GlianaAI

GlianaAI lists a configurable /x402/audio/speech route and dedicated eleven-v3 and tts-1.5-max synthesis routes. Speech prices start at $0.014. The configurable route says the body selects a model and the price applies to the exact request body; its default does not establish the cost of a different model or a longer request. Keep the selected model and voice in the job record. The provider's wider endpoint list also includes language, image and video tools, so a cheaper unrelated route cannot price a speech call. A model name in the listing is the provider's advertised routing choice, not evidence of pronunciation quality, accent coverage or equivalent output across providers.

3 speech-synthesis endpoints of 29 · $0.014/call (POST /x402/audio/speech) · Base/SOL · uptime 30d 100% · 212 ms · 27 buyers in 30 days

Voix Off TTS FR — French neural voiceovers pay-per-call

Voix Off TTS FR lists /generate for a short French IVR message, at $0.03. The listed endpoint describes a 15-second use case. Its broader description also discusses other voiceover formats; the captured endpoint list establishes the IVR route used here. For a telephone greeting, check the destination system's codec, sample rate and channel requirements before buying a larger script. Listen to business names, dates and phone numbers in a representative sample. A French-language positioning claim does not establish suitability for a particular regional accent, telephone system or caller experience without inspecting the resulting audio.

1 speech-synthesis endpoint of 1 · $0.03/call (GET /generate) · Base · uptime 30d 99.9% · 940 ms · 0 buyers in 30 days

kino402

kino402 describes /v1/audio/tts as ElevenLabs speech. The recorded reference quote is $0.007; the provider says charging depends on characters. Keep that reference separate from an estimate for the exact script. This gateway also lists video and other generation tools, which are excluded from the synthesis type. For a batch workflow, inspect how the provider reports a failed request and whether the client can recover the delivered asset before retrying. Repeated synthesis of identical text can still require a new payment. A local cache keyed by text, model, voice and settings can make a repeated spoken notification more predictable to operate.

1 speech-synthesis endpoint of 10 · $0.007/call (POST /v1/audio/tts) · Base/SOL · uptime 30d 100% · 252 ms · 13 buyers in 30 days

What the calls cost at different volumes

The median selected endpoint costs $0.012 per call. At that same request shape, 100,000 calls would cost $1,200.00. This multiplication is a comparison aid, not a provider invoice or a promise that an arbitrary request will use the reference quote.

Each service at the price of its main speech-synthesis endpoint, the one named in the table above.

Budget for useful results as well as submissions. A retry, an additional research step or a separate product can add another charge; no-match and partial-failure handling depend on the particular endpoint. For a variable body, model or input size, inspect the returned payment requirement for that exact request. Keep the quoted route and paid unit in the job record rather than projecting from the cheapest unrelated tool on a gateway.

How an agent reaches the HTTP 402 payment step

An unpaid request lets the caller inspect a payment requirement. The example below pairs the documented route with a real HTTP 402 payload previously captured by x402-list's monitor. Its capture date belongs to that evidence; it is not a claim that a paid result was tested today. The documented route has a reference quote of $0.001; inspect the challenge for the exact request terms.

curl -i 'https://voice.forgemesh.io/v1/tts/base?text=Hello&voice=F1'

HTTP/1.1 402 Payment Required

{
  "x402Version": 2,
  "resource": {
    "url": "https://voice.forgemesh.io/v1/tts/base"
  },
  "accepts": [
    {
      "asset": "0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913",
      "extra": {
        "name": "USD Coin",
        "version": "2"
      },
      "payTo": "0x962794FB8aD92941fAf1ab446b9e30637B86F331",
      "amount": "1000",
      "scheme": "exact",
      "network": "eip155:8453",
      "maxTimeoutSeconds": 1800
    }
  ]
}

402 of ForgeMesh Voice for GET /v1/tts/base, as stored by the monitor (terms unchanged since 2026-09-14, last check 2026-10-26 17:50 UTC), with every accepted payment exactly as stored and without error, extensions, resource.mimeType and resource.description, which are not payment terms: the full body is also on the service page, under Payment evidence. Amounts are in the asset's smallest unit: USDC has 6 decimals, so 1000 is $0.001.

Inspect the request schema before adapting the example. A documented POST endpoint needs the corresponding JSON inputs; a GET endpoint needs valid query values. Keep the payment requirement tied to the intended resource, network, asset and request. The captured challenge shows what payment access looks like. It does not prove that a paid call will return accurate output, complete a background job or satisfy the application's success criteria.

What this comparison measures

The listing requires an online, payment-ready service, a usable description, an advertised price and at least 90% measured uptime over 30 days. Task selection then reads its endpoint list and keeps the matching routes. Those listing checks establish a discovery threshold, not a certification of output accuracy.

Uptime is based on monitoring checks; average latency measures those checks. Buyer counts are distinct paying wallets observed for the service in the measurement window, with the limits described on the service page. A successful unpaid HTTP 402 response confirms a payment challenge rather than successful delivery after settlement.

Treat the provider's output contract and our monitoring as complementary evidence. Test a representative input, inspect source dates and missing states, and define a successful result before putting a recurring workflow on the endpoint. For comparison details, open the linked service page and the site's measurement methodology. Recheck the exact payment requirement when the input or advertised product changes.

Questions

How to pay

Each service answers an unpaid call with HTTP 402 and its price; your agent signs a USDC payment on the listed network and retries. No API key, no account, no subscription. See what x402 is.