Best Text-to-Speech APIs for AI Apps

As of October 13, 2026, no text-to-speech API leads on listening, delay, and price together. On Audio Realism, Cartesia Sonic 3.6 has the top win rate among the APIs priced here: 72.3% over 209 battles.

ElevenLabs' Eleven v3 Conversational follows at 63.3% over 461 battles. Google's Gemini 2.5 Pro TTS Preview and OpenAI's GPT-4o mini TTS trail, at 49.4% and 40.4%.

Vendors supply the delay claims, billing units differ, and rights depend on the plan. This guide keeps the three apart.

It covers the speech layer of the AI App Stack. For the model behind a call, see Best LLMs for Voice AI Agents.

Three questions need three kinds of evidence

Naturalness, delay, and rights each have their own source, and none of the sources answers the other two.

Table 1 Download CSV

Question	Evidence	What it can't tell you
Which voice sounds most human?	Audio Realism: expert listeners hear two same-script clips and pick the one that "sounds more human"	Delay, price, or rights
How soon does audio start?	Vendor inference figures, such as ElevenLabs' roughly 75 ms for Flash, plus your own timing	What a caller hears, which includes network and playback
Can I ship it?	Pricing pages and API docs, read October 13, 2026	Whether the voice sounds good

The managed-platform phone ranking on our voice benchmarks page times whole platforms, including their speech recognition and model choices. It isn't a TTS-model ranking, so this guide doesn't use it.

The listening shortlist has small samples at the top

Audio Realism fits listener votes with a Bradley-Terry model and reports an Elo-style rating. We re-read the arena's methodology page on October 13. It tests one female and one male voice per model, chosen as the most conversational American-English voice, on mostly short prompts of up to about 40 words. Its scored set is 500 private prompts. It states no minimum number of votes or uncertainty, and its API rows carry none.

Table 2 Download CSV

Rank	Model	Win rate	Battles	Elo
1	Freya TTS Adam V1	79.2%	379	1433
2	Bland Speech v3	77.9%	620	1377
3	Kalpa TTS Beta v0.1	74.4%	540	1364
4	Cartesia Sonic 3.6	72.3%	209	1339
6	ElevenLabs Eleven v3 Conversational	63.3%	461	1232
7	MAI-Voice-2	61.3%	609	1213
8	Cartesia Sonic 3.5	61.9%	801	1210
13	Gemini 2.5 Pro TTS Preview	49.4%	644	1112
19	GPT-4o mini TTS	40.4%	607	1032
20	Gemini 3.1 Flash TTS Preview	37.4%	645	1031
23	ElevenLabs Eleven v3	30.0%	636	968

Source: Design Arena leaderboard, category audiorealism, last updated October 11, 2026, 29 rows and 6,662 votes, with selected rows shown. Rank is by Elo, so a higher win rate can sit lower.

Three readings follow from the table. Freya, Bland, and Kalpa sit within 69 Elo points, and with no published uncertainty we can't call the order of the first two settled. Cartesia Sonic 3.6 has 209 battles, and we flag any row under 300 battles as thin. That flag is our editorial line, not the arena's, and 11 of the 29 rows fall below it. Last, a family name doesn't carry the result: Eleven v3 and Eleven v3 Conversational are 264 Elo points apart.

Seven rows are unmapped to our model pages, including Deepgram Flux TTS, Eleven v4, and Eleven v4 Turbo. Eleven v4 and v4 Turbo each sit under 300 battles, so we leave them out of the table rather than rank them. Amazon Polly, Google's Chirp 3 HD, and Kokoro aren't on this board at all, which means they're unmeasured here, not poor. We haven't priced or reviewed API access for the listening leaders, so read their own terms before shortlisting one.

Each pick has a job and a losing case

Table 3 Download CSV

Pick	Fits when	Documented surface	Loses when
Cartesia Sonic 3.6	You stream audio into a phone or voice-agent loop	WebSocket contexts with a cancel request for interruptions, 8 kHz mu-law and A-law output, 44 languages, dated snapshot IDs such as sonic-3.6-2026-08-27	You need broad language coverage beyond 44, or want certainty from a sample above 300 battles
ElevenLabs	One vendor should cover voices, cloning, and agents	Flash v2.5 (32 languages, about 75 ms inference), v3 (70+ languages), instant and professional cloning, mu-law and A-law at 8 kHz	Per-character cost matters, or you'd pick plain Eleven v3, which placed 23rd
Google Cloud TTS	Your estate already runs on Google Cloud	Chirp 3 HD with 53 language entries and streaming, though not SSML in streaming requests	Listening quality is the deciding factor, since its Gemini voices rank 13th and lower here
Amazon Polly	You need IVR, read-aloud, or speech marks inside AWS	Four engines, mp3, Ogg, PCM, mu-law, A-law, and JSON speech marks	You need long requests, since one call allows 6,000 characters of which 3,000 can be billed
OpenAI GPT-4o mini TTS	Your stack already bills through OpenAI	Token pricing and a 2,000-token input cap	Naturalness leads the brief, at a 40.4% win rate over 607 battles
Kokoro	You want local control and Apache-2.0 weights	82 million parameters, nine language codes, 24 kHz output	No engineer will own inference, scaling, and quality testing

Where each pick loses decides more than where it wins. Cartesia's alias sonic-3.6 points at the latest stable snapshot and updates on its own, so pin the dated ID if production behavior must not change. ElevenLabs says professional clones add per-generation latency over default and instant clones. Google's streaming path drops SSML, which matters if your scripts depend on pronunciation tags.

Billing units differ, so a budget needs one stated assumption

Per-character, per-credit, per-token, and per-call-minute charges are different units, and none converts cleanly into another. Take one hypothetical month: 10 million characters of speech, about 12,500 minutes at 800 characters a minute, which is Cartesia's own 750-to-800-credits-per-minute conversion. The table counts character or audio charges only. Taxes, text-input tokens, cloning, concurrency add-ons, and any voice-agent platform fee are excluded.

Table 4 Download CSV

Route	Unit and list rate	Free allowance	Cost for the month
Google Standard or WaveNet	$4 per 1M characters	4M a month	$24
Google Neural2	$16 per 1M characters	1M a month	$144
Google Chirp 3 HD	$30 per 1M characters	1M a month	$270
Amazon Polly Standard	$4 per 1M characters	5M a month	$20
Amazon Polly Neural	$16 per 1M characters	1M a month for 12 months	$144, or $160 after year one
Amazon Polly Generative	$30 per 1M characters	100K a month for 12 months	$297, or $300 after year one
Cartesia Scale plan	1 credit per character, 8M credits for $299, then $38 per 1M	None beyond the plan	$375
ElevenLabs Flash	$0.04 per 1,000 characters	Not stated	$400
ElevenLabs v3	$0.08 per 1,000 characters	Not stated	$800
Gemini 2.5 Flash TTS	$10 per 1M audio tokens at 25 tokens a second	None	$187.50
Gemini 3.8 Flash TTS (Preview)	$9 per 1M audio tokens through December 31, 2026	None	$168.75, then $337.50 from January 1, 2027
OpenAI GPT-4o mini TTS	$12 per 1M audio tokens	Not stated	Not computed

Sources: Google Cloud, Amazon Polly, Cartesia, ElevenLabs, and OpenAI, read October 13, 2026.

OpenAI's page gives no per-minute estimate, so convert from the audio tokens your own scripts produce. Google's pricing page doubles Gemini 3.8 Flash TTS on January 1, 2027, so follow model changes if you budget past that date. Cost beside the listening table is a trade you make, not a ranking: the cheapest rows here are also the ones the arena doesn't measure.

Commercial rights follow the plan and the voice

Cartesia's pricing page lists a commercial use license from the $5 Pro plan up and none on Free. Instant voice cloning starts at Pro, and professional voice clones come with two slots on Startup and four on Scale. ElevenLabs' pricing page lists a commercial license from Starter up and none for Free. Its pricing calculator sizes the plan.

We haven't reviewed the cloud providers' service terms or per-voice restrictions for this guide. A selected voice, engine, or plan can carry its own limits, so open the terms for the exact option before you ship, and don't infer rights from a family name.

A matched-script trial needs about 200 votes per pair

Keep published results, vendor claims, and your own trial apart in your notes. The trial borrows Audio Realism's structure: same text for every vendor, one female and one male voice each, blind pairs, and randomized order.

Use 40 scripts, 10 from each class, in the languages you serve, checked against the multilingual leaderboard for the model side:

Table 5 Download CSV

Script class	What to listen for
Names, addresses, order codes	Pronunciation and correction
Numbers, URLs, units	Text normalization
Long support answer	Pacing and drift
Short confirmations and interrupted sentences	Cutoff and restart

Five listeners across 40 scripts gives 200 votes per pair. At 200 votes, a 50% win rate carries about plus or minus 6.9 points at 95% confidence, so a 56-to-44 split proves nothing. Four hundred votes narrows that to 4.9 points. This is a simple binomial estimate, not the arena's Bradley-Terry fit.

For delay, time from your application's request to the first audible sample with the real player, from the regions where your callers sit. ElevenLabs notes that network round trips add 20 to 200 ms and a 500 ms player buffer is common, which is why its 75 ms inference figure never matches what a caller hears. Cartesia's chunks carry a step_time field, a server-side timing you can compare against your own clock.

Log one row per request:

Table 6 Download CSV

Field	Example
Script and class	names-07
Vendor, model, snapshot	cartesia sonic-3.6-2026-08-27
Voice ID and region	voice-f1, us-east
Request to first audio (ms)	612
Format and sample rate	pcm_mulaw, 8,000 Hz
Listener votes	31 of 40
Rights check	Plan and license page recorded

Reject a candidate when it mispronounces a required entity twice in one script class, when its p95 exceeds your budget, when the rights don't fit, or when the monthly cost at your volume is over budget. Watch Cartesia Sonic 3.6's battle count next: it needs 91 more battles to clear our 300-battle flag.