Best AI Voice Generators in 2026: 10 Providers Compared

Updated August 2026 · Every fact verified against official vendor documentation on August 11, 2026.

Decide faster: Compare pricing · Find your fit · How we verify

Every price, licensing term, and feature claim on this page was read from official vendor documentation on August 11, 2026, with the source linked beside the fact. Our standardized audio benchmark is designed but not yet running — this page contains no listening-test claims and no scores.

The ranking at a glance

The order below is an editorial judgment built from documented pricing, licensing clarity, and feature coverage for these text to speech (TTS) products — not from listening tests, which we have not run yet. How we chose explains the basis; each provider name links to its full assessment.

Not sure which fits you? Jump to who should choose what for recommendations by use case, see how the numbers were built in pricing, normalized, or read how we chose first.

How we chose — and what we do not claim

Every changing fact on this page — every price, model name, quota, free-tier limit, and licensing term for these TTS products — was read from the vendor's own current documentation on August 11, 2026, and is cited in the same sentence or table cell where it appears. Licensing language is quoted from the operative terms wherever practical, and rights verified for one plan are never extended to another. The full protocol is on our methodology page; the writing rules are in our editorial policy.

Just as important is what this page does not claim. We have not generated audio from any of these products, created accounts with them, or run listening tests. Our standardized audio benchmark — the same short prompts generated under the same conditions by every provider — exists as a finished design, not a running system (methodology explains it). Until it runs, this page carries no naturalness verdicts, no "we tested" claims, and no scores of any kind. Where a vendor makes a quality or latency claim, it is labeled as the vendor's claim.

The ranking order is therefore an editorial judgment about documented fit: how many kinds of buyer each product serves well on verified pricing, licensing clarity, and feature coverage — with ties broken toward the provider whose terms answer the questions a buyer must ask. A provider whose terms are silent on commercial use, or whose pricing cannot be found on any public page, ranks below one that documents both, regardless of how its voices might sound. When the audio benchmark runs, this page will be refreshed with measured comparisons, and the change log will say so.

Where official sources disagree with each other — and several do — the disagreement is stated openly rather than silently averaged; see what we could not verify. One familiar name is absent from the list entirely: PlayHT. The reason is in the FAQ.

Independence note: this page contains no affiliate links, and no vendor paid for placement or influenced the order. How the site is funded is set out in how we make money; factual errors are handled under our corrections policy.

Pricing, normalized — and how to read it

AI voice generator pricing splits into two families that resist direct comparison. Creator-facing tools sell subscriptions denominated in credits, "voice generation time", or points; API vendors bill per character. Only some vendors publish a conversion between their unit and characters — where one exists, the table normalizes to USD per 1,000 characters; where the vendor bills by time with no character conversion, the row is normalized per minute and says so; where no conversion exists at all, the cell says "not normalizable" rather than inventing one.

The spread is enormous: verified list prices run from $0.004 per 1,000 characters (Google Standard/WaveNet; Amazon Polly Standard, both checked Aug 11, 2026) to $0.160 per 1,000 characters (Google Studio voices, checked Aug 11, 2026) — a 40× range before creator subscriptions, whose plan-effective rates run higher still. Three reading rules:

• Formulas are shown in the cells. Every normalized figure is computed from the vendor's own listed price and quota, so you can check the arithmetic.
• List prices are not measured costs. These are the vendors' published rates as of August 11, 2026 — not what we paid, because we have not generated anything. When our benchmark runs, actual generation costs will be published alongside.
• Compare within a row's billing basis only. A per-minute figure and a per-character figure are not comparable to each other without assumptions about speaking rate that the vendors do not publish.

Feature matrix: API, cloning, languages, formats, licensing

The matrix below compares what each vendor documents, not what we have measured. Two columns deserve special care: language counts are the vendors' own claims — several vendors publish different counts on different pages, and the cells say so — and the commercial-use column covers paid tiers only; free-tier rights are in the pricing table above, because they are usually different and usually worse.

ElevenLabs · expressive control at a premium

ElevenLabs is the most complete creator package in this comparison: the deepest documented expressive controls, self-serve voice cloning, and a commercial license that starts on the $6/month Starter plan (elevenlabs.io/pricing, checked Aug 11, 2026). The price reality: its API lists at $0.10 per 1,000 characters for Multilingual models (elevenlabs.io/pricing/api, checked Aug 11, 2026), and plan-effective rates run higher still. The bluntest con: it is the premium-priced option at every level — and its own two pricing pages disagree by roughly 1.65×–2×, depending on plan, on what a subscription includes.

What the documentation supports

• A model for every job. Eleven v3 covers 70+ languages and takes inline audio tags such as [whispers] and [laughs] (5,000-character cap); Eleven Multilingual v2 (29 languages) is the model the vendor calls most stable for long-form (10,000-character cap); Eleven Flash v2.5 (32 languages) is the speed tier, with a vendor-claimed ~75 ms model latency, a 40,000-character cap, and half the per-character API price (subscription credit rate: 0.5–1 credit/char, exact rate unpublished) (elevenlabs.io/docs/models, checked Aug 11, 2026).
• Clean commercial licensing from the bottom paid tier. The commercial license is listed from Starter at $6/month (pricing, checked Aug 11, 2026), and the ToS states the user retains rights in the output (§4(c), updated Mar 31, 2026) (terms, checked Aug 11, 2026).
• Self-serve cloning at consumer prices. Instant voice cloning is listed from Starter, and Professional Voice Cloning from Creator ($22/month) (pricing; cloning docs, both checked Aug 11, 2026).
• A real developer surface. REST and WebSocket APIs, Python and Node SDKs, and a per-request cost header for billing transparency (API docs, checked Aug 11, 2026). Unused subscription credits roll over for up to two months (billing docs), capped at 2× the monthly quota (elevenlabs.io/pricing, both checked Aug 11, 2026).

The trade-offs

Cost, and cost transparency. Plan-effective rates work out to $0.165–$0.200 per 1,000 characters on the Multilingual basis (computed in the pricing table from elevenlabs.io/pricing, checked Aug 11, 2026) — roughly 41×–50× Google's and Amazon's $0.004 standard tiers, and about 11× Azure's $0.015 pay-as-you-go rate. Worse for budgeting: the main pricing page quotes plans in credits while the API pricing page quotes the same plans in characters at roughly 1.65×–2× the credit figures depending on plan, and the vendor does not reconcile the two (pricing vs API pricing, both checked Aug 11, 2026). Output quality is also tier-gated: 192 kbps MP3 requires Creator or above, and 44.1 kHz PCM/WAV requires Pro or above (API reference, checked Aug 11, 2026).

Free tier and commercial use

The free plan (10,000 credits/month) is strictly non-commercial and requires attribution (ToS §1(c), updated Mar 31, 2026; terms and billing docs, both checked Aug 11, 2026). Anything you intend to publish commercially needs at least Starter. We could not verify whether the free plan requires a card, or the exact attribution wording — the vendor's help article on it was inaccessible at check time (see what we could not verify).

Voice cloning and consent

Instant cloning is listed from the Starter plan; Professional Voice Cloning is restricted to your own voice, enforced by microphone verification and a 24-hour wait, and the ToS requires that any voice you clone be yours or one you are expressly authorized to share (cloning docs; terms, both checked Aug 11, 2026).

Who should look elsewhere

Cost-sensitive, high-volume API users: Google Cloud and Amazon Polly's standard tiers list at $0.004 per 1,000 characters (Google pricing; Polly pricing, both checked Aug 11, 2026) — 25× cheaper than ElevenLabs' Multilingual API rate. Anyone planning to publish work from a free tier, and anyone who needs to clone a voice other than their own through the professional route, should also pass. Instead consider Google Cloud Text-to-Speech for volume economics.

Google Cloud Text-to-Speech · the verified price floor

Google Cloud Text-to-Speech is the cheapest verified option in this comparison — $0.004 per 1,000 characters on the Standard and WaveNet tiers — and pairs it with the largest free allowance: 4 million characters per month on those tiers, always free (cloud.google.com/text-to-speech/pricing, checked Aug 11, 2026). The price reality: it is pay-as-you-go infrastructure with no plans and no studio. The bluntest con: even free usage requires a billing account, overage auto-bills, and voice cloning sits behind a sales allow-list.

What the documentation supports

• The lowest list price and biggest free bands we verified. Standard and WaveNet voices at $4 per 1M characters, with 4M free characters/month each; Chirp 3 HD, Studio, Neural2, and Polyglot voices carry 1M-character free bands (pricing; free-tier docs, both checked Aug 11, 2026).
• A current model line, including prompt-steerable TTS. Gemini-TTS models (gemini-2.5-flash-tts, gemini-2.5-pro-tts, gemini-3.1-flash-tts-preview) accept natural-language style prompts and multi-speaker input; Chirp 3: HD is the GA streaming tier with 53 locales; the vendor claims 380+ voices across 75+ languages overall (product page and docs, checked Aug 11, 2026).
• Serious automation surface. REST and gRPC, client libraries, bidirectional streaming, published per-project quotas, and async long-audio synthesis up to 1M input bytes (API reference, checked Aug 11, 2026).
• A no-training commitment. Google's service terms state it will not train on customer data without permission (Service Specific Terms §18, cloud.google.com/terms/service-terms, checked Aug 11, 2026).

The trade-offs

The newest models break the price comparison: Gemini-TTS is token-billed, and the only vendor conversion published (25 audio tokens per second) normalizes output to about $0.015/minute on 2.5 Flash and $0.030/minute on 2.5 Pro — with text-input tokens on top that cannot be converted at all (pricing, checked Aug 11, 2026). Studio voices cost $0.160 per 1,000 characters — 40× the Standard tier and above ElevenLabs' API list rate (same source). The terms ban competitive use and training models on the output, and TTS is not on Google's list of generative services covered by its output indemnity (service terms, checked Aug 11, 2026). Whether TTS falls under Google's separate "Generative AI Services" terms is genuinely ambiguous in the published documents — we say so rather than guess (see what we could not verify).

Free tier and commercial use

The free bands are generous but not casual: "You must enable billing to use Text-to-Speech", usage beyond the band auto-bills, and the $300/90-day trial requires a payment method (pricing; free-tier docs, both checked Aug 11, 2026). On commercial use, the terms draw no free-vs-paid distinction: all usage runs under the same B2B contract, in which the customer retains IP in Customer Data (ToS §5.1, cloud.google.com/terms, checked Aug 11, 2026).

Voice cloning and consent

Chirp 3: Instant Custom Voice is the cloning route, and it is gated: access is granted through a sales allow-list, a fixed per-language consent script must be recorded, reference audio is capped at 10 seconds, and synthesis with a custom voice costs $60 per 1M characters (ICV docs, checked Aug 11, 2026). There is no self-serve path.

Who should look elsewhere

Creators who want a browser studio, plan pricing, or self-serve cloning — none of which exist here. Buyers who need output IP indemnity for generated speech, and AI-adjacent companies constrained by the competitive-use ban, should read the service terms closely before committing. Instead consider Murf AI for an editor-first workflow with quotable commercial rights.

Murf AI · studio workflow, quotable commercial rights

Murf AI is the studio pick: a timeline-style editor sold as Studio plans from $29/month, next to a separately-priced API at $0.03 per 1,000 characters (murf.ai/pricing; API plans, both checked Aug 11, 2026) — and the cleanest commercial-use sentence in this comparison: "You can use Murf created voices for commercial purposes" (ToS §5.1, terms, checked Aug 11, 2026). The bluntest con: the free trial produces nothing you can keep — no downloads, self-evaluation only — and Studio time-billing cannot be compared per-character with rivals.

What the documentation supports

• Commercial rights without fine print. ToS §5.1's grant is tier-unqualified for paid plans; the stated limits are no resale of the voices themselves and no using output to train AI (ToS, Dec 19, 2024, checked Aug 11, 2026).
• A cheap, simple API. Pay-as-you-go at $0.30 per 10,000-character block ($0.03/1k) with a $2 minimum purchase, an official Python SDK, published rate limits, and a 100,000-character free API trial (help.murf.ai; API docs, both checked Aug 11, 2026).
• A real-time model line. Falcon is Murf's streaming model with a vendor-claimed sub-130 ms time to first audio (murf.ai/api, checked Aug 11, 2026); GEN2 is the studio-quality generation model (API docs, checked Aug 11, 2026).
• Broad export options on paid Studio plans: MP3, WAV, FLAC, A-law, and µ-law at 8/24/48 kHz (help.murf.ai), plus a fuller format set on the API (API reference, both checked Aug 11, 2026).

The trade-offs

Studio bills Voice Generation Time — hours of output — with no published character conversion, so its $0.133–$0.242/minute normalized range cannot be compared per-character against rivals; the vendor's own approximation is about 6 minutes of VGT per 1,000 words (help.murf.ai, checked Aug 11, 2026). Procurement teams should also know the public pricing page is a JavaScript-only shell: we verified prices from Murf's own production pricing endpoint, whose payload carries a "pricingTestGroup":"B" flag — meaning served prices may be A/B-test variants (murf.ai/pricing data, checked Aug 11, 2026). Murf's own pages also disagree on catalog size — 150 to 300+ voices and 33 to 35+ languages depending on the page (murf.ai and help center, checked Aug 11, 2026). One dated warning for API users: the legacy Gen2 streaming model deprecates on August 16, 2026 — five days after this page's verification date — and users must move to Falcon 2 (murf.ai/api/docs, checked Aug 11, 2026).

Free tier and commercial use

The Studio free trial gives 10 minutes of VGT with no card required — but no downloads of any kind, and the ToS limits trial use to "self-evaluation" (§10.7) (help.murf.ai; ToS, both checked Aug 11, 2026). It answers "should I subscribe?", not "can I use this output?". The API trial is more useful: 100,000 characters with one API key (help.murf.ai, checked Aug 11, 2026).

Voice cloning and consent

Cloning is enterprise-only and not self-serve: roughly 1–2 hours of studio-quality recordings, a 1–4 week turnaround, and explicit written consent from the voice owner required by ToS §4.2 (help.murf.ai; terms, both checked Aug 11, 2026).

Who should look elsewhere

Anyone who wants self-serve voice cloning — Murf has none at any self-serve tier. Anyone who evaluates tools through free output will find the trial useless by design. And buyers who need vendor-stable public pricing pages for procurement should note the A/B-tested, JavaScript-only pricing surface. Instead consider ElevenLabs, where instant cloning is listed from the $6/month tier.

Speechify · the cheapest API plans, with a licensing trap

Speechify's developer platform is the cheapest entry-level character-billed subscription here: the API Starter plan is $10/month for 1 million characters — $0.010 per 1,000 — falling to $0.006 per 1,000 in Scale-tier overage, a rate that ties Azure's largest commitment tier (speechify.ai/pricing, checked Aug 11, 2026). The price reality: "Speechify" is three products with three different sets of terms. The bluntest con: the $29/month Premium app most consumers find first is expressly non-commercial (ToS §4.4, speechify.com/terms, checked Aug 11, 2026), and the API's commercial rights rest on inference, not on a quotable grant.

What the documentation supports

• The cheapest entry-level subscription-class API rates we verified. $10/month for 1M characters, with overage at $10/$8/$6 per additional 1M depending on tier — the $6/1M Scale overage ties Azure's largest commitment tier (speechify.ai/pricing, checked Aug 11, 2026).
• A genuinely capped free API tier. 50,000 characters/month with a hard cap — usage pauses instead of billing you (speechify.ai/pricing, checked Aug 11, 2026).
• Current streaming-first models. simba-3.2 is English-only and streaming-native with the vendor's lowest claimed time-to-first-byte; simba-3.0 covers six languages; simba-multilingual claims 30+ (docs.speechify.ai, checked Aug 11, 2026).
• Studio paid plans carry an explicit rights statement: commercial usage rights and "you own the audio output" (speechify.com/pricing-studio, checked Aug 11, 2026).

The trade-offs

The product split does real damage to clarity. The consumer app's Premium plan is metered in words (a 150,000-word/month contractual floor, with 1M words/month guaranteed through 2026 — usage limits, checked Aug 11, 2026) and its terms list "commercial use" among the grounds for restricting an account (speechify.com/usage-limits, checked Aug 11, 2026; consistent with the non-commercial rule in ToS §4.4, terms, checked Aug 11, 2026). The API's supplemental terms designate output as "Customer Content" and impose a duty to disclose AI generation (§1.4) — but contain no express commercial-use grant to quote (API terms, Jun 10, 2024, checked Aug 11, 2026). Studio displays only annual prices, and its download formats are unpublished (pricing-studio, checked Aug 11, 2026).

Free tier and commercial use

Three free tiers, three answers. The API free tier (50K chars/month) sits under terms that are silent on free-vs-paid commercial rights — we state that silence rather than fill it (API terms, checked Aug 11, 2026). Studio's free plan states outright: "No commercial usage rights", and no voice cloning (pricing-studio, checked Aug 11, 2026). The app's free tier offers 10 "robotic sounding voices" and is non-commercial like Premium (pricing, checked Aug 11, 2026).

Voice cloning and consent

Studio cloning requires a paid plan and a 20-second sample. API cloning takes a 10–30 second sample plus a mandatory consent record containing the speaker's full name and email, and prohibits cloning minors, deceased persons, and well-known political figures (cloning API docs; API terms, both checked Aug 11, 2026).

Who should look elsewhere

Anyone about to buy Premium expecting to publish commercial work — that is the single easiest licensing mistake in this comparison. API customers whose legal review requires an express commercial grant in writing, and multilingual API projects needing current-generation quality beyond simba-3.0's six languages, should also look on. Instead consider Murf AI, whose paid-tier commercial grant is one quotable sentence.

Amazon Polly · lowest-cost infrastructure with the cleanest ownership

Amazon Polly matches the cheapest verified rate in this comparison — $0.004 per 1,000 characters on the Standard engine, pay-as-you-go with no plans (aws.amazon.com/polly/pricing, checked Aug 11, 2026) — and pairs it with the cleanest ownership language: "Your Polly output belongs to you", with "no restrictions on storing and reusing generated speech" (Polly FAQ, checked Aug 11, 2026). The bluntest con: no voice cloning, no creator UI — and AWS's own free-tier pages contradict each other.

What the documentation supports

• Four engines, four prices. Standard is $0.004 per 1,000 characters; the other three engines run to $0.100/1k — the full ladder is in the pricing table (pricing, checked Aug 11, 2026).
• Ownership without tiers. The output-ownership statements are not tier-scoped; the one stated restriction is a ban on using output to train competing AI models (Service Terms §50.5, service terms, Jul 29, 2026, checked Aug 11, 2026).
• Telephony-grade plumbing. MP3, Ogg, PCM, µ-law/A-law output, Speech Marks JSON for lip-sync and highlighting, streaming responses, async batch to 100K billed characters, and a per-request billed-character response header (API reference, checked Aug 11, 2026).
• Free replay. Cached and replayed audio costs nothing — you pay once per synthesis (pricing, checked Aug 11, 2026).

The trade-offs

The newest quality tier comes with caveats AWS itself documents: the Generative engine is available in only 9 regions, produces no Speech Marks, and carries a vendor-acknowledged hallucination risk with an "emergency stop mechanism" that "reduces but does not eliminate the risk" (developer guide, checked Aug 11, 2026). By default AWS may store and use processed text to improve its AI services, with opt-out only via an AWS Organizations policy — worth knowing for confidential scripts (Service Terms §50.3, service terms, checked Aug 11, 2026). And there is no self-serve cloning of any kind (features page, checked Aug 11, 2026).

Free tier and commercial use

The free tier is two-track and ambiguous. Per-engine monthly character bands exist (Standard 5M; Neural 1M; Long-Form 500K; Generative 100K — the latter three "for the first 12 months", while the Standard band's duration is stated inconsistently across AWS's own pages), but accounts opened on or after July 15, 2025 fall under a $200-credit regime whose precedence over the character bands AWS does not state (pricing; free-tier FAQ, both checked Aug 11, 2026). A card is required at signup regardless. On commercial use, no tier-scoped restriction appears anywhere — free-band output runs under the same service terms as paid (service terms, checked Aug 11, 2026).

Voice cloning and consent

There is no self-service cloning. Brand Voice — a custom voice built with the Polly team from actor recordings — is a negotiated enterprise engagement with no published price or timeline, and AWS's Responsible AI Policy forbids depicting a person's voice without consent (features; policy, both checked Aug 11, 2026).

Who should look elsewhere

Anyone who wants voice cloning, a consumer editor, or output beyond the AWS console-and-SDK workflow. Teams outside the 9 Generative-engine regions who want its newest tier, and testers unwilling to hand AWS a card just to experiment, should also pass. Instead consider ElevenLabs for cloning and an editor, or Google Cloud Text-to-Speech for the same price floor with bigger free bands.

Azure Speech in Foundry Tools · the biggest catalog, behind enterprise gates

Azure Speech in Foundry Tools — Microsoft's renamed Azure AI Speech (Learn TTS overview, checked Aug 11, 2026) — claims the largest catalog among the cloud platforms in this comparison: "more than 500" prebuilt voices (HD voices comparison, checked Aug 11, 2026) across 100+ languages and locales (Learn TTS overview, same date). The price reality: $0.015 per 1,000 characters pay-as-you-go, falling to $0.006 on the largest commitment tier (Retail Prices API, checked Aug 11, 2026). The bluntest con: free-tier output sits outside the commercial-use grant, and every voice-cloning route requires being "managed by Microsoft".

What the documentation supports

• Scale pricing that rewards commitment. Standard neural runs $0.015 per 1,000 characters pay-as-you-go, falling to $0.006/1k on the largest commitment tier — the full tier ladder is in the pricing table (prices.azure.com, corroborated by Learn docs, both checked Aug 11, 2026).
• Catalog depth and expressive tiers. DragonHD voices add automatic emotion detection and roughly 60 styles; personal voice speaks 91 languages (TTS overview and model docs, checked Aug 11, 2026).
• Enterprise plumbing. REST plus Speech SDKs in C#, Python, JavaScript, and Java; batch synthesis; on-prem containers; the widest output-format list here (including AMR-WB and G.722 for telephony); and a machine-readable price API (REST reference, checked Aug 11, 2026).
• Clear data terms. Output is Customer Data, Microsoft claims no ownership of it, custom voices are exclusive to the customer, and using output to train competing products is banned (Product Terms, checked Aug 11, 2026).

The trade-offs

Transparency, oddly, is the weak spot for a company this size: the canonical pricing page rendered every dollar figure as "$-" placeholders at check time, so our figures come from Microsoft's official Retail Prices API (eastus region — prices vary by region), corroborated by the Learn documentation's $15/1M rate (prices.azure.com; Learn quotas doc, both checked Aug 11, 2026). The newest models — MAI-Voice-2 and MAI-Voice-2-Flash — are public preview, carry no SLA, and Microsoft says they are not recommended for production (model docs via the TTS overview, checked Aug 11, 2026). Two billing quirks: every CJK character counts as two billed characters (same overview), and signup requires a non-prepaid card for identity verification even for free use (account page, checked Aug 11, 2026).

Free tier and commercial use

The F0 tier gives 0.5M neural characters per month at no cost — but read the grant: commercial use of prebuilt neural voice output is licensed "For Customers of the paid tier TTS Service only", which leaves F0 output outside the quoted permission (Microsoft Product Terms, checked Aug 11, 2026). Treat F0 as an evaluation tier, not a production one.

Voice cloning and consent

All three cloning routes — personal voice (1-minute sample), professional voice (30 minutes to 3 hours of recordings), and the MAI instant clone preview (5–60 seconds) — are Limited Access features: "only customers managed by Microsoft" are eligible, recorded talent consent is required and may be verified biometrically, and AI-disclosure duties to end users are mandatory (Limited Access page, checked Aug 11, 2026).

Who should look elsewhere

Hobbyists planning to monetize on the free tier — the grant does not cover them. Anyone who wants self-serve cloning without an enterprise relationship, and teams that need the newest voice models under a production SLA, should also pass. Instead consider Cartesia, where a commercial license and instant cloning are self-serve at the lowest verified entry price.

OpenAI · steerable speech with no free way in

OpenAI's Audio API is the developer's shortcut to instruction-steerable speech: gpt-4o-mini-tts takes plain-language direction over accent, emotion, tone, and even whispering (TTS guide, checked Aug 11, 2026). The price reality: legacy tts-1 lists at $0.015 and tts-1-hd at $0.030 per 1,000 characters, but the current flagship model is token-billed with no published conversion (pricing, checked Aug 11, 2026). The bluntest con: there is no free way to try it, and no way to budget its flagship model per character from official documentation.

What the documentation supports

• Instruction steering. gpt-4o-mini-tts (snapshots 2025-12-15 and 2025-03-20) accepts natural-language delivery instructions — a control surface the character-billed incumbents document differently or not at all (TTS guide, checked Aug 11, 2026).
• The cleanest ownership assignment in this comparison. "Customer… owns all Output. OpenAI hereby assigns to Customer all OpenAI's right, title, and interest, if any, in and to Output" (Services Agreement §4.1, eff. Jan 1, 2026, business terms, checked Aug 11, 2026).
• Predictable pricing on the legacy models. $15/1M and $30/1M characters for tts-1 and tts-1-hd respectively (pricing, checked Aug 11, 2026).
• Standard developer conveniences: mp3, opus, aac, flac, wav, and pcm output, chunked streaming, and Python/JavaScript SDKs (API reference, checked Aug 11, 2026).

The trade-offs

The catalog is small and English-leaning: 13 built-in voices (9 on the legacy models), described by the vendor as "optimized for English", with 57 languages claimed via Whisper lineage (TTS guide, checked Aug 11, 2026). Requests cap at 4,096 characters (2,000 tokens for gpt-4o-mini-tts), forcing chunking for long-form (API reference, checked Aug 11, 2026). gpt-4o-mini-tts bills $0.60/1M text-input tokens and $12/1M audio-output tokens with no token-to-character or token-to-minute conversion published, so its real-world cost cannot be normalized (pricing, checked Aug 11, 2026). There is no browser studio product at all. Apps must disclose to end users that the voice is AI-generated (TTS guide, checked Aug 11, 2026).

Free tier and commercial use

No free TTS tier of any kind exists; API use requires prepaid credits purchased in advance, which expire after one year and are non-refundable — a card is effectively required before any use (prepaid billing help, checked Aug 11, 2026). ChatGPT's free voice feature is not a workaround: its output is expressly non-commercial and may not be repackaged as standalone audio files (Service Terms §8, updated Jun 12, 2026, service terms, checked Aug 11, 2026). Paid API output, by contrast, is owned by the customer outright under the §4.1 assignment quoted above.

Voice cloning and consent

Custom voices arrived via the API: a consent recording of a fixed OpenAI-specified phrase is required first, samples are capped at 30 seconds, and organizations may hold at most 20 voices (TTS guide, checked Aug 11, 2026). One honesty note: the documentation references a "Text-to-Speech Supplemental Agreement" that we could not locate anywhere — cloning-rights specifics may live in a document we could not read (see what we could not verify).

Who should look elsewhere

Anyone who needs a free tier, a browser UI, or a large voice library. Long-form producers unwilling to chunk at 4,096 characters, and buyers who must budget the flagship model's cost precisely in advance, should also pass. Instead consider Google Cloud Text-to-Speech, which pairs steerable Gemini-TTS models with free bands and a published character price floor.

Cartesia · budget real-time voice with self-serve cloning

Cartesia has the cheapest commercial entry point we verified: the $5/month Pro plan includes 100,000 credits, a commercial-use license, and instant voice cloning from a clip of 10 seconds or less (cartesia.ai/pricing, checked Aug 11, 2026). The price reality: a published 1-credit-per-character conversion makes its rates fully comparable — $0.037–$0.050 per 1,000 characters by plan (same source). The bluntest con: by default your inputs and outputs may train Cartesia's models unless you file an opt-out, and the governing terms predate the plans they govern.

What the documentation supports

• Commercial rights at pocket money prices. The "commercial use license" is listed from Pro at $5/month, and the ToS states Cartesia claims no ownership of outputs (§5.3) (pricing; terms, both checked Aug 11, 2026).
• Honest unit math. 1 credit = 1 character for standard TTS, published in the vendor's own pricing FAQ — which is rare among credit-billed vendors and makes every plan normalizable (pricing, checked Aug 11, 2026).
• Real-time infrastructure. Vendor-claimed sub-90 ms time to first audio, SSE and WebSocket streaming, and published per-plan concurrency limits (TTS: 2/3/5/15 by plan) (docs; pricing, both checked Aug 11, 2026).
• A documented deprecation calendar. Current models are sonic-3.5 (stable snapshot sonic-3.5-2026-05-04) and sonic-3; the remaining sonic-2 and sonic-turbo snapshots (and the older sonic-3-2025-10-27 snapshot) sunset on October 20, 2026 — dated in the docs, not discovered by surprise (model docs, checked Aug 11, 2026). 42 supported languages are listed with their full code list (same page).

The trade-offs

The data-governance default is the big one: the ToS states that inputs, outputs, and interactions "may be used by Cartesia to train… its models", with opt-out by request form (§5.3(c)–(d), terms, checked Aug 11, 2026). The same ToS is dated June 14, 2024 — before the current plan names existed — so the commercial-tier mapping rests on the pricing page rather than the contract (terms; pricing, both checked Aug 11, 2026). Format support is also asymmetric: MP3 and WAV containers are available only on the Bytes endpoint, while the SSE and WebSocket streaming endpoints return raw audio only (format docs, checked Aug 11, 2026).

Free tier and commercial use

The free tier gives 20,000 credits (20,000 characters) per month and is expressly non-commercial: the ToS restricts free use to "personal, non-commercial use only" (§4.1, §5.3(b), terms, checked Aug 11, 2026). Whether a card is required to sign up is not stated (pricing, checked Aug 11, 2026).

Voice cloning and consent

Instant cloning from a ≤10-second clip is included from Pro; the higher-fidelity Pro Voice Clone requires at least 30 minutes of audio, a Startup plan or above ($49/month), a one-time 1M-credit training fee, and bills output at 1.5 credits per character (cloning docs; pricing, both checked Aug 11, 2026). The ToS bars making any person's voice available without express permission (§4.2(13)) — but no consent-verification workflow is documented in the cloning docs, so enforcement rests on the contract, not the product (terms, checked Aug 11, 2026).

Who should look elsewhere

Teams with strict data-governance requirements who cannot rely on an opt-out form, MP3-dependent workflows built on streaming endpoints, and anyone who needs consent-verified cloning as a product feature rather than a contract clause. Instead consider Azure Speech in Foundry Tools, whose enterprise terms make data handling and voice exclusivity explicit.

MiniMax · cheap, capable, and legally split

MiniMax's speech API is aggressively priced: speech-2.8-turbo at $0.06 and speech-2.8-hd at $0.10 per 1,000 characters pay-as-you-go, with Rapid Voice Cloning at $1.50 per voice — the lowest documented cloning fee in this comparison (platform.minimax.io pricing, checked Aug 11, 2026). The bluntest con: the licensing surface is split and partly silent — the consumer app's output is not clearly commercially usable even on paid plans, and the API terms grant ownership without ever saying "commercial use permitted".

What the documentation supports

• Expressive current models. Speech-2.8, released January 23, 2026 per the vendor's release notes (release notes, checked Aug 11, 2026), adds 19 native interjection tags — (laughs), (sighs), (gasps) and the like (API reference, checked Aug 11, 2026).
• A big documented catalog. 332 system voices on the official voice-ID list and 40 named language_boost options (voice list; API FAQ, both checked Aug 11, 2026).
• The largest long-form capacity here. The async API accepts up to 1,000,000 characters in a single request (API reference, checked Aug 11, 2026).
• Billing transparency at the request level: responses carry billable-character counts, and output options include word- and sentence-level timestamps (API reference, checked Aug 11, 2026).

The trade-offs

Subscriptions are denominated in "audio points" with no published points-to-characters conversion, so their value cannot be cost-compared — only pay-as-you-go can (subscription pricing, checked Aug 11, 2026). No free API speech tier is documented, so there is no documented way to evaluate before paying (pricing, checked Aug 11, 2026). The API contract is with Nanonoble Pte. Ltd. under the platform ToS effective March 30, 2026, which requires deep-synthesis disclosure — services built on it must add identifiers and "a prominent mark" telling the public deep synthesis was used (platform ToS, checked Aug 11, 2026).

Free tier and commercial use

The consumer MiniMax Audio app offers "daily time-limited free credits" with an unpublished quota, but its ToS permits "personal, non-commercial use only" (consumer ToS, Jun 19, 2025; paid service terms, updated Feb 12, 2026, both checked Aug 11, 2026) — and the paid consumer plans' commercial position is ambiguous. On the API side, the platform ToS states the customer retains ownership of generated content, with no non-commercial restriction — but also no affirmative commercial-use grant to quote (platform ToS, checked Aug 11, 2026). Commercial users belong on the API, and cautious ones should ask MiniMax for the grant in writing.

Voice cloning and consent

Rapid Voice Cloning costs $1.50 per voice plus synthesis-priced previews, from a 10-second to 5-minute sample (pricing, checked Aug 11, 2026). Cloning is gated on account verification, unused cloned voices are deleted after 7 days, and the data-subject consent duty sits with the customer — no consent-capture flow is documented in the product (cloning guide; platform ToS, both checked Aug 11, 2026).

Who should look elsewhere

Consumer-app users who need commercial rights, buyers who need a documented free tier to evaluate against, and procurement teams that require an express commercial grant or convertible subscription units. Instead consider OpenAI, whose terms assign output ownership in one unambiguous sentence.

Resemble AI · security-first voice, price on request

Resemble AI is the security-and-consent specialist of this list: every generated file carries the PerTh perceptual watermark by vendor statement, consent tooling is built into professional cloning, and the MIT-licensed Chatterbox family is the standout self-hosting route in this comparison (start-here page; GitHub, both checked Aug 11, 2026). The bluntest con: you cannot learn what its hosted voice generation costs from any public page, and its terms never say who owns the audio or whether you may sell it.

What the documentation supports

• Watermarking by default. The vendor states the PerTh watermark is auto-embedded in generated audio, hosted and open-source alike (start-here page, checked Aug 11, 2026).
• A genuine open-source escape hatch. Chatterbox is MIT-licensed — Turbo (350M parameters), Nano (110M, on-device), and Multilingual V3 (23+ languages) — usable without an API key via Hugging Face (GitHub, checked Aug 11, 2026).
• Deep delivery control on the hosted API: SSML with temperature, exaggeration, a deterministic seed, inline tags such as [laugh] and <whisper>, plus grapheme/phoneme timestamps, streaming, and webhooks (API reference, checked Aug 11, 2026).
• A no-expiry free entry point. The Flex plan is $0/month with pay-per-use credits that never expire and no card required at sign-up (pricing, checked Aug 11, 2026).

The trade-offs

Pricing opacity is unmatched in this comparison — in the wrong direction. The hosted API's only current TTS model is resemble-ultra; every previous hosted model, including all hosted Chatterbox variants, is end-of-life and "can no longer generate audio" (model-versions doc, checked Aug 11, 2026). Yet no voice-generation rate is published anywhere: the public pricing page lists per-second rates only for the deepfake-detection products, and the app's pricing URL returned a 404 at check time (resemble.ai/pricing, checked Aug 11, 2026). Plans are listed at Team $350/month and Business $1,000/month with no synthesis rate attached (same page). The company's public packaging now centers deepfake detection and AI security; voice generation remains fully documented but visibly de-emphasized (pricing; product page, both checked Aug 11, 2026).

Free tier and commercial use

Flex is $0/month with no card and never-expiring credits, but the free voice-generation quota is unpublished (pricing, checked Aug 11, 2026). More importantly, the ToS contains no output-ownership assignment and no commercial-use grant or restriction at any tier — a silence a buyer must resolve with sales before committing; the one clear prohibition is using output to train or improve other products (ToS, last update Jul 30, 2024, checked Aug 11, 2026). Self-hosted Chatterbox is different: MIT-licensed, so its permissions are in the license itself (GitHub, checked Aug 11, 2026).

Voice cloning and consent

Rapid Clone builds a voice from 10 seconds to 3 minutes of audio in under a minute of training; Professional Clone takes 10–25+ minutes of audio and, per the vendor's FAQ, requires "explicit verifiable consent" with consent workflows "built into the platform" — though how Rapid Clone verifies consent is ambiguous in the docs (cloning docs; product page, both checked Aug 11, 2026). Note the gate: the Voice Cloning API requires the Business plan at $1,000/month list (same sources).

Who should look elsewhere

Anyone who needs transparent self-serve pricing, cloning-API access below a $1,000/month plan, or explicit output-rights language their legal team can quote. Instead consider Cartesia, which publishes every rate and sells a commercial license from $5/month.

Who should choose what

These recommendations are derived from the documented facts above — pricing, licensing, and feature coverage as of August 11, 2026 — not from listening tests. Each links back to the full assessment.

Video and content creators

ElevenLabs is the default: expressive tags, self-serve cloning, and a commercial license from $6/month (pricing, checked Aug 11, 2026) — if the per-character premium fits your volume. Teams that live in a timeline editor should weigh Murf AI, whose paid plans pair the workflow with the plainest commercial-use sentence in this comparison. Speechify Studio paid plans state output ownership explicitly, but display annual prices only.

Developers and API builders

For volume, Google Cloud Text-to-Speech ($0.004/1k characters and 4M free characters/month on Standard/WaveNet — pricing, checked Aug 11, 2026) and Amazon Polly (same floor, cleanest ownership terms) are the anchors. For subscription budgeting, Speechify's API is the cheapest character-billed plan class we verified; Murf AI's API at $0.03/1k with a $2 minimum (API plans, checked Aug 11, 2026) is the low-friction middle. Teams already on OpenAI infrastructure get instruction steering and clean ownership from OpenAI — priced per character only on its legacy models.

Enterprise and compliance-led buyers

Azure Speech in Foundry Tools is built for this buyer: commitment tiers to $0.006/1k characters, on-prem containers, customer-exclusive custom voices, and explicit data terms (Retail Prices API; Product Terms, both checked Aug 11, 2026). Google Cloud offers comparable contractual footing with a stated no-training commitment. Where audio provenance itself is the requirement, Resemble AI's default watermarking is the documented differentiator.

Real-time agents and conversational apps

Latency numbers here are all vendor claims — none are our measurements. Cartesia claims sub-90 ms to first audio and publishes per-plan concurrency (pricing, checked Aug 11, 2026); ElevenLabs claims ~75 ms model latency on Flash v2.5 (models page, checked Aug 11, 2026); Murf AI claims sub-130 ms time to first audio on Falcon (murf.ai/api, checked Aug 11, 2026). All three ship WebSocket streaming; MiniMax's turbo tier is the budget entrant. When our benchmark runs, this is one of the first claims it will test.

Budget and free usage

The largest documented free allowances are Google Cloud's 4M characters/month (Standard/WaveNet) and Amazon Polly's 5M Standard characters/month band (Google pricing; Polly pricing, both checked Aug 11, 2026) — both under B2B terms with no tier-scoped commercial restriction found, but both requiring billing setup. Speechify's API free tier (50K chars/month) is the rare hard-capped one. The caution: most creator-tool free tiers are expressly non-commercial — see the FAQ. The cheapest paid commercial licenses we verified are Cartesia at $5/month and ElevenLabs at $6/month (Cartesia pricing; ElevenLabs pricing, both checked Aug 11, 2026).

Self-hosted and on-device

One documented route exists in this comparison: Resemble AI's MIT-licensed Chatterbox family — Turbo (350M), Nano (110M, on-device), and Multilingual V3 (23+ languages), usable without an API key (GitHub, checked Aug 11, 2026). Azure Speech offers on-prem containers, but as a licensed cloud service rather than open source.

What we could not verify

Honesty about gaps beats false completeness. The items below are questions our source records explicitly mark as unverifiable from official documentation as of August 11, 2026 — the page states them rather than guessing. Where a vendor's own pages disagree with each other, both figures are attributed to their pages above, never averaged.

• ElevenLabs: whether the free plan requires a card; the exact required attribution wording (the help-center article returned an error); the exact Flash credit-per-character subscription rate; any reconciliation of the ~1.65×–2× disagreement between its credits-based and characters-based pricing pages; and the EEA terms — every ToS citation on this page is to the non-EEA ToS, and an EEA variant exists that we did not review.
• Murf AI: what the pricing page actually renders to a browser (it is a JavaScript-only shell; we used Murf's production pricing endpoint); whether served prices are A/B-test variants (the payload flags a test group); exact voice and language counts (Murf's pages disagree); Enterprise and Murf Dub pricing; and the API trial's card requirement.
• Speechify: the Premium annual price (not rendered — we print none); card requirements across all three products; the complete API audio-format list; Studio download formats and monthly prices; and who owns consumer-app output (the ToS is silent).
• OpenAI: the referenced "Text-to-Speech Supplemental Agreement" (not locatable anywhere we searched); any token-to-character conversion for gpt-4o-mini-tts; custom-voice pricing; and whether new accounts receive trial credits.
• Google Cloud: whether TTS falls under Google's "Generative AI Services" terms; and the total voice count (a second official blurb still says 220+ voices/40+ languages against the product page's 380+/75+).
• Amazon Polly: which free-tier regime (character bands vs $200 credits) governs accounts opened on or after July 15, 2025; the Standard band's duration (AWS pages state it inconsistently); and Brand Voice pricing, timeline, and consent procedure.
• Azure Speech: every dollar figure on the canonical pricing page (rendered as "$-"; we used Microsoft's official Retail Prices API instead); MAI-Voice list pricing; and whether HD/MAI preview voices draw down the free F0 allowance.
• Cartesia: whether the free plan requires a card, and whether any consent-verification step exists in the cloning UI (none is documented).
• MiniMax: any audio-points-to-characters conversion (subscriptions are therefore not cost-compared here); the consumer app's free quota; and the paid consumer plans' commercial-use position.
• Resemble AI: the hosted TTS list price (published nowhere); the free voice-generation quota; commercial-use and output-ownership rights (the ToS is silent); and how Rapid Clone verifies consent.

Frequently asked questions

Change log

Published August 11, 2026 · all prices and terms read from official sources on this date.