Amazon Polly public-data review: paying more buys less

The rate card

All four rates are per million characters of input text, read off the official pricing page today. The per-minute and per-hour columns apply the 1,000 characters per minute equivalence this site uses across every per-character vendor, which is ElevenLabs' own published figure rather than a CartSignal assumption.

Two housekeeping notes that matter more than they look. AWS bills input characters, and the quota documentation states that SSML tags are not counted as billed characters — so pronunciation fixes, pauses, emphasis and rate changes are free on this meter, which is not true of vendors whose emotional direction is written inline in the script. And the pricing page states you can "cache and replay Amazon Polly's generated speech at no additional cost", so a fixed prompt or a static page read aloud is a one-off charge rather than a recurring one.

Finding 1: price and coverage run in opposite directions

This is the finding that should decide the engine, and it is invisible on the pricing page because the pricing page lists only prices. Cross-referencing the rate card against the voice list and each engine's own documentation gives this:

Above the $16.00 Neural tier, every dollar you add subtracts reach. Generative drops 13 language variants and 9 regions for an 87.5% price rise. Long-form, at 6.25 times the Neural rate, is four en-US voices and two es-ES voices, available in us-east-1 and nowhere else — AWS's long-form page lists exactly one region and then the line "Other regions not available".

The counts above are CartSignal's, taken row by row from the AWS voice table, and three of them can be checked against AWS's own stated totals: AWS says it "currently offers 43 voices in a generative variant", that neural voices "are available in 36 languages and language variants", and that there are "four en-US and two es-ES long-form voices". All three match. AWS publishes no total for the Standard engine, so 60 voices across 29 variants is our count alone.

The consequence for data residency

If your processing has to stay inside a jurisdiction, the engine choice is made for you before quality is considered. Long-form is unavailable outside North Virginia at any price. Generative runs in ten regions including Frankfurt, London and Zurich — but AWS adds that London, Canada (Central) and Zurich carry only 22 of the 43 generative voices. Neural is the broadest premium option at 19 regions, and the only one listed in GovCloud. This interacts directly with the marking duties in our EU AI Act labeling guide: an EU deployment that wants both the best-sounding engine and local processing has a much shorter voice list than the marketing count suggests.

Finding 2: the dearest engine is beaten by the one below it on everything except timing

Set Long-form ($100.00) against Generative ($30.00) directly and the comparison is one-sided in the wrong direction. Long-form costs 3.3 times more and offers fewer voices (6 against 43), fewer languages (2 against 23) and fewer regions (1 against 10). It is also the slower of the two to schedule in bulk: both are capped at 1 transaction per second on the asynchronous API.

There is exactly one thing Long-form has that Generative does not, and it is not sound: speech marks. AWS's generative documentation states that "Support for generating speech marks is currently not available", and the speech marks page confirms it from the other side — "Speechmarks are available when using neural, long-form, or standard text-to-speech engines." Generative is the one engine missing from that list.

So the honest statement of what the $100.00 tier sells is narrow: it is the premium-expressiveness engine for applications that also need word-level timing, in English or Spanish, in one region. Everywhere else, public data does not support paying 3.3 times Generative's rate. AWS's own quality numbers, below, do not rescue it either.

Finding 3: word timing doubles the whole rate card

Speech marks are the metadata that tells you when each word and each mouth shape happens. AWS's own examples for them are "synchronize speech with facial animation (lip-syncing)" and "highlight written words as they're spoken" — which is to say, read-along apps, captions and avatar lip sync. Two published rules combine into a price nobody quotes:

• The documentation states that when you request speech marks, "Amazon Polly returns this metadata instead of synthesized speech". One request gives you one or the other.
• The pricing page states that "you are charged based on the number of characters of text that you convert either to speech or to Speech Marks metadata".

Audio plus timings is therefore two passes over the same characters and two charges. The effective rate card for any job that needs both:

The doubling also eats the free tier, because the allowance wording is "5 million characters per month for speech or Speech Marks requests" — one pool, both purposes. And it removes the cheap-and-expressive corner of the card entirely: a talking-head product that wants the best Polly voice and lip sync has to buy Long-form at $200.00 per million effective, which is the point at which Polly stops being the cheap option in this directory at all. An hour of that is $12.00, the same as an hour of narration on an ElevenLabs Starter subscription.

Finding 4: AWS publishes error rates, and they get worse as the price rises

Almost no vendor in this directory publishes quality measurements of its own output. AWS does, in the AI Service Card for Amazon Polly, which states it applies to releases current as of 11 February 2026. Placed next to the prices:

Read the two right-hand columns together and the trade is explicit. Moving from Neural to Generative costs 87.5% more and buys 6.96% higher satisfaction and 2.3 times the critical error rate. Moving from Neural to Long-form costs 525% more for 5.91% higher satisfaction and about twice the error rate. The cheapest of the three measured engines is the most reliable one by a wide margin.

AWS defines both metrics on the same page. Voice QA is the proportion of 1,000 to 2,000 test prompts per locale on which trained listeners flagged a critical error, drawn from nine categories including cutoffs, audio glitches, hallucinations, mispronunciation and inconsistent speaker identity. Customer satisfaction is an average 1-to-7 rating from at least 50 native listeners across 15 samples. These are AWS's own measurements on AWS's own test sets, not an independent benchmark, and a 0.4-point gap on a 7-point average is not a small perceptual difference just because it is a small percentage. But the direction is what matters for buying: on this card, extra money buys expressiveness and buys back risk.

The service card's narrative text gives a second, older-looking set of figures — Standard and Neural voices "scoring a voice CSAT of 5.5" with "a critical pronunciation error rate of less than 5%", and Generative at 5.8 with "critical pronunciation errors less than 3%". Those do not match the per-engine table on the same page, and AWS does not date either set, so the table is used above and the discrepancy is noted rather than reconciled.

Why the premium engines err more

AWS is unusually candid about the mechanism, in a note on the generative documentation that is worth reading before shipping generative speech into anything unattended. It describes an "imposed emergency stop mechanism" that halts synthesis when hallucination is detected, warns that the model "might end up cutting a word during a generation step, thus rendering half the word", and closes: "This safety feature reduces but does not eliminate the risk. In some edge cases, the model might continue generating random content beyond the stop threshold. Do not assume complete protection against hallucinated speech output."

Both the generative and long-form pages carry a second warning with a direct content consequence: model updates "could result in slight variations to the way the voices sound", and AWS names the affected case itself — "use cases with different content parts synthesized over a long time period — for example, a season of podcasts". If you are producing serial audio in one voice over months, the premium engines are the ones that may drift under you.

One more wrinkle on the Neural engine, from its own page: "For certain voices, Amazon Polly might automatically switch to the Standard engine when accuracy is critical. For example, this can happen during letter-by-letter spelling." You can therefore hear an audible quality drop mid-sentence on an engine you are paying four times the Standard rate for. AWS does not state which rate those characters are billed at.

Finding 5: the free tier is a supply, not an audition

Most free tiers in this directory are demonstrations you cannot publish. Polly's is different in two ways, and both are structural rather than generous.

The licence does not change. There is no free plan with its own terms — the free tier is a billing allowance on the same service — so output made inside it carries the same rights as output you pay for. The Polly FAQ states: "As between you and AWS, your Polly output belongs to you. If you input text into Polly that belongs to a third party, we require that you have the rights to do so." Compare that with the rule our commercial use rights guide found at ElevenLabs, where content created outside a paid subscription cannot be used commercially and paying later does not license it retroactively. On Polly there is nothing to upgrade.

The Standard allowance is large and carries no stated expiry. The pricing page attaches "for the first 12 months" to Neural, Long-form and Generative, and attaches no duration at all to the 5 million Standard characters a month. At 1,000 characters a minute that is roughly 83 hours of speech a month, against 10 minutes a month on an ElevenLabs free account. We are not calling it perpetual: AWS's free-tier catalogue page, which is where an offer is labelled always-free or 12-months-free, renders client-side and returned no Polly entry on fetch, so the absence of a duration on the pricing page is the only evidence either way.

Measured in list-price value the free tier inverts along with everything else on this card: the Long-form allowance is worth $50 a month at list, Standard $20, Neural $16 and Generative just $3. The dearest engine hands out the most valuable free sample.

A related point for anyone with a privacy review to pass: the AI Service Card states that "AWS does not store nor use inputs or outputs generated through the Amazon Polly service to train Amazon Polly service" and that Polly "is not storing inputs nor outputs" beyond pronunciation lexicons. That applies at every tier, free included — unlike the billing-dependent data terms our commercial-rights guide found on Google's Gemini API.

Finding 6: six languages exist only at $4.00, and thirteen cannot reach it

The Standard engine is usually described as the legacy option. For part of Polly's language list it is the only option. Working through the voice table, six language variants have no neural, generative or long-form voice at all — Arabic (arb), Icelandic, Romanian, Russian, Welsh and Welsh English. For those, the $4.00 rate is not a budget decision; it is the entire offer, and the concatenative Standard engine is what your listeners get.

It runs the other way too. Thirteen language variants have no Standard voice, including Catalan, Czech, Cantonese, Finnish, Belgian French, Austrian and Swiss German, Gulf Arabic and four English variants (Irish, New Zealand, Singaporean, South African). In those, the cheapest available rate is $16.00, not $4.00 — so the "50x cheaper than ElevenLabs" headline this site publishes in its cost-per-character ranking is a statement about 29 of Polly's 42 language variants, not all of them.

One structural detail worth knowing before planning an upgrade path: every language with a generative voice also has a neural voice, with no exceptions, while 13 neural languages have no generative option. Generative is a strict subset. Moving up from Neural never gains you a language.

Throughput, request size and the limits that bite first

Polly's quotas matter more than usual here because there is no plan to upgrade — the limits are per engine and per account, and the expensive engines are the slow ones.

Two readings follow. Real time is where Standard is untouchable at ten times the request rate of every other engine, which is exactly the shape of an IVR or an announcement system. Batch is where Generative and Long-form fall off a cliff at 1 transaction per second — a back catalogue of ten thousand articles is a queue measured in hours before any synthesis time is counted, and these quotas are adjustable on request, which is a support ticket rather than a price.

The request-size limits are the other thing to design around, and they are generous in an unexpected place:

• Real time: up to 3,000 billed characters, 6,000 total. The gap between the two numbers is the SSML allowance, which is unbilled — so half your payload can be markup for free. Output audio is capped at 10 minutes per request, after which "any remaining speech is cut off".
• Batch: up to 100,000 billed characters, 200,000 total — about 100 minutes of speech in one job, written to your own S3 bucket.
• Lexicons: 100 per account, 40,000 characters each, up to five applied per request. This is the free way to fix a brand name or a technical term permanently, and it is the feature that most directly substitutes for the re-generation cycles that cost money on per-character vendors.
• SSML gaps: the audio, lexicon, lookup and voice tags are not supported, break is capped at 10 seconds, and prosody rate will not go below -80%.

Voice cloning: not sold, at any price

Polly is the clearest counterexample to the taxonomy our voice cloning cost guide built, where vendors ration clones by slot and gate them by consent mechanism. Polly rations nothing because it sells nothing self-serve. Its custom-voice product, Brand Voice, is described as "a custom engagement where you work with the Amazon Polly team to build a Neural Text-to-Speech (NTTS) voice for the exclusive use of your organization", in which AWS helps "identify the persona, identify an actor or actress and record their speech" before training the model. The AI Service Card is blunter: "Customers cannot create brand voice on their own through Polly."

The route is an AWS account manager. No price, no turnaround time and no minimum recording length is published. That is a real constraint rather than a pricing gap — if cloning a specific person's voice is why you are shopping, Polly is not a cheaper way to do it, it is a different purchase entirely, and the shortlist is ElevenLabs, Cartesia or Speechify. It also means the consent question that dominates the rest of that page does not arise here in product form: there is no upload box to attest against.

What Amazon Polly actually sells

• Text to speech over an API and a console, in four engines, billed per input character with no plan or minimum.
• Speech marks — sentence, word, viseme and SSML timing metadata — on the Standard, Neural and Long-form engines, billed as a separate pass at the same per-character rate.
• Asynchronous synthesis of up to 100,000 billed characters per job, delivered to your own S3 bucket.
• Bidirectional streaming on the Generative engine, announced on 20 March 2026 alongside 10 new generative voices, letting text stream in while audio streams out for live agents and game characters.
• Pronunciation lexicons and SSML, both unbilled, for permanent pronunciation control.
• Brand Voice, a custom voice built by AWS under a sales engagement, with no published price.

And what it does not sell, which is most of what the rest of this directory is: there is no studio or timeline editor, no dubbing or translation product, no avatars, no voice agent bundle with an LLM in it, no transcription, and no subscription of any kind. Polly is a component. If you want a finished workflow rather than an endpoint, the comparison to read is what a minute of AI voice actually costs, by job.

Best-fit use cases

• High-volume programmatic narration where the per-character rate is the whole decision — Polly Standard is the floor of this site's cost-per-character ranking
• IVR, announcements and other real-time prompts, where Standard's 80 transactions per second is ten times any other engine's
• Accessibility and read-along features, where speech marks are the product and the doubled rate is still under a cent a minute on Standard
• Teams already inside AWS, where billing, IAM, PrivateLink, CloudWatch and S3 delivery are already solved
• Anything needing Arabic, Icelandic, Romanian, Russian or Welsh, where Standard is the only engine offered
• Long-running evaluation, since the Standard free allowance is roughly 83 hours of commercially usable speech a month

Limitations to verify

• The Generative engine cannot produce speech marks. If you need lip sync or word highlighting, the best-sounding engine is unavailable at any price
• Speech marks double the bill, because audio and timings are separate billed passes over the same characters, and they share the free allowance too
• Long-form runs in us-east-1 only, in two languages, at 3.3x the Generative rate
• The premium engines are the error-prone ones on AWS's own testing, and AWS explicitly warns not to assume complete protection against hallucinated generative output
• Generative and Long-form batch synthesis is capped at 1 transaction per second, which is the binding limit on back-catalogue work
• Voice cloning is not self-serve. Brand Voice is a sales engagement with no published price, turnaround or data requirement
• No content filtering on input. The service card states Polly "does not filter inappropriate or sensitive input text" and puts moderation on the customer
• The Standard free tier's duration is unconfirmed. The pricing page states 12 months for the other three engines and no duration for Standard; AWS's free-tier catalogue page renders client-side and could not be read
• Neural may silently fall back to Standard for accuracy-critical passages such as letter-by-letter spelling, with no statement of which rate applies
• Polly publishes no statement located during this review about machine-readable marking of its generated audio, which matters for the duties in our EU AI Act labeling guide

How it fits against other tools here

Polly is the cheap floor of this directory and it is a component rather than a product, which is the same distinction that separates the developer APIs from the studio subscriptions everywhere on this site. At $4.00 per million characters Standard is about 50 times below an ElevenLabs Starter subscription's effective $200.00, and at $16.00 Neural still undercuts every subscription in the category. But compared like for like at the premium end — Polly Long-form and ElevenLabs v3 both at $100.00 per million, Polly Generative and Deepgram Aura-2 and the OpenAI tts-1-hd model all at $30.00 — the field narrows to a band of roughly 3.3x, and the 50x gap is two different products rather than two prices. The full ranking with every vendor's basis stated is on ElevenLabs alternatives, ranked by cost per character.

Where Polly is genuinely different is the absence of packaging. There is no plan ladder that stalls, no tier that charges more per character than the one beneath it, no credits to expire and no annual commitment to negotiate — the traps this site has documented at ElevenLabs, Speechify, Rask and Deepgram all require a subscription to exist in. What replaces them is an engine choice that changes your language coverage, your region coverage, your error rate and your access to timing data all at once, which is a harder decision than picking a tier and a less obvious one. The category overview is AI voice tools, and the job-by-job costing is best AI voice tools.

FAQ

How much does Amazon Polly cost?

Per character of input text, with no plan and no minimum. As of 20 September 2026: Standard $4.00 per million characters, Neural $16.00, Generative $30.00, Long-form $100.00. At 1,000 characters a minute that is $0.24, $0.96, $1.80 and $6.00 for an hour of finished speech. GovCloud (US) lists Standard at $4.80 and Neural at $19.20, a flat 20% premium. Standard is the cheapest published rate in this directory, about 50x below an ElevenLabs Starter subscription's effective $200.00 per million.

Which Amazon Polly engine should I use?

Start from coverage and reliability, because on Polly they do not track price. Neural at $16.00 reaches 36 language variants in 19 regions, supports all speech marks, and has the lowest critical error rate of the three measured engines at 0.77%. Generative at $30.00 scores highest on AWS's satisfaction measure (6.15 against 5.75 of 7) but errs on 1.79% of prompts, reaches 23 variants in 10 regions and produces no speech marks. Long-form at $100.00 is 3.3x Generative for 6 voices in 2 languages in one region. Standard at $4.00 is the broadest, the fastest, and for six languages the only engine offered.

Why does Amazon Polly cost double for lip sync or word highlighting?

Because the timing data is a second billed pass. Polly returns speech marks "instead of synthesized speech", and the pricing page charges for characters converted "either to speech or to Speech Marks metadata" — so audio plus timings is two requests and two charges. Effective rates become $8.00 per million on Standard, $32.00 on Neural and $200.00 on Long-form. On Generative it cannot be bought at all: AWS states that support for generating speech marks is currently not available there.

Is the Amazon Polly free tier usable for commercial work?

Yes. There is no free plan with separate terms — the free tier is a billing allowance on the same service — so free output carries identical rights. The Polly FAQ states "As between you and AWS, your Polly output belongs to you", with the condition that you hold the rights to any third-party input text. That is the reverse of ElevenLabs, where our commercial rights guide found the licence attaches to the moment of generation, so paying later does not cover work already made. The Standard allowance is 5M characters a month, roughly 83 hours of speech, with no duration stated on the pricing page.

Can you clone a voice with Amazon Polly?

Not yourself. Brand Voice is "a custom engagement where you work with the Amazon Polly team" to build a neural voice for your organisation, including identifying an actor and recording them, and the AWS AI Service Card states that "Customers cannot create brand voice on their own through Polly." The route is an AWS account manager, and no price, turnaround or minimum recording length is published. So Polly sits outside the slot-and-consent model our voice cloning cost guide found everywhere else.

Has CartSignal tested Amazon Polly?

No. This is a public-data review built from the official Polly pricing page, the developer guide's quota, voice list, neural, generative and long-form pages, the Polly FAQs, the AWS AI Service Card for Amazon Polly and the AWS announcement of 20 March 2026 — all read on 20 September 2026. No audio was generated, so nothing here evaluates how any voice sounds. The quality figures are AWS's own published measurements on its own test sets, reproduced as vendor statements rather than independent benchmarks. Every per-minute rate, percentage and multiple is CartSignal arithmetic over AWS's published prices and counts.

Source links

• Official Amazon Polly pricing (all four per-million rates, GovCloud rates, free tier amounts and durations, Speech Marks billing, caching) — read 20 September 2026
• Amazon Polly quotas (transactions per second per engine, 3,000 and 100,000 billed-character limits, "SSML tags are not counted as billed characters", 10-minute audio cap, lexicon limits)
• Amazon Polly voice list (per-engine availability for every voice and language variant, Brand Voice note)
• Amazon Polly generative voices (43 voices, 10 regions, the 22-voice region subset, no speech marks, the hallucination stop mechanism)
• Amazon Polly long-form voices (six voices in two languages, "Other regions not available")
• Amazon Polly neural voices (36 language variants, 19 regions, newscaster style, automatic fallback to the Standard engine)
• Amazon Polly speech marks (returned instead of audio; available on neural, long-form and standard engines)
• AWS AI Service Card for Amazon Polly (per-engine Voice QA and CSAT figures and methodology, no training on inputs or outputs, no input content filtering, Brand Voice is not self-serve)
• Amazon Polly FAQs ("As between you and AWS, your Polly output belongs to you", no restriction on stored or replayed prompts)
• Amazon Polly features (Brand Voice custom engagement description)
• AWS announcement, 20 March 2026 (10 new generative voices across 8 locales, London and Canada regions, Bidirectional Streaming API)