What Is Amazon Polly? Pricing per Character, Free Tier, Every Voice and Engine, SSML, Python Code and Commercial Use
Amazon Polly is AWS's text-to-speech service: you send it text (up to 3,000 characters per request, or 100,000 in a background task), pick one of more than 100 voices in 42 languages and variants, and get back an MP3, OGG or PCM audio file in a second or two. It has four voice engines with four prices per million characters: Standard $4, Neural $16, Generative $30 (the most human-sounding) and Long-form $100 (for audiobooks and articles). One million characters is about 23 hours of speech, so a typical 6,500-character news article costs about 3 cents with a standard voice, 10 cents neural and 20 cents generative. Accounts on the classic AWS Free Tier get 5 million standard characters a month, and for the first 12 months 1 million neural, 500,000 long-form and 100,000 generative characters a month; newer accounts use Free Tier credits instead. The audio belongs to you, and you can store and replay it as often as you like at no extra charge, which is the single biggest cost lever. Below: what Polly is, the four engines compared, every English voice, pricing with worked examples, the console, Python and CLI, SSML and pronunciation lexicons, speech marks for lip sync and highlighting, real-time streaming for voice assistants, a "listen to this article" setup, Polly vs other text-to-speech options, commercial use, limits and error fixes.
The furniture store's phone menu had been recorded by Jake six years earlier, in one take, with the delivery van idling in the background. Every time opening hours changed, the menu was wrong until someone found a quiet room and Jake's patience. The owner also wanted an audio version of the sofa care guide for the website and narration for three short assembly videos. Jake offered to record all of it again. Ethan typed one sentence into the Amazon Polly console, pressed Listen, and Jake went very quiet.
Building something with Polly? The cards below cover Lambda, S3 access, CloudFront, the voices built into Windows, and billing alerts, the pieces you will most often need around it.
What is Amazon Polly?
Text-to-speech (TTS) turns written text into spoken audio. Your phone reads messages aloud with it, GPS apps give directions with it, and screen readers use it to read the web to people who cannot see the screen. Amazon Polly is AWS's managed version: an API (plus a console page for trying it) that takes text and returns natural-sounding speech, with nothing to install, no model to host and no GPU to pay for.
People use Polly for:
- Phone systems and contact centers: menu prompts, hold messages and dynamic answers ("your order ships on Friday"); Amazon Connect and Amazon Lex use Polly voices.
- Articles and blogs: "listen to this article" buttons and podcast versions of written content.
- Video narration: explainer videos, product demos, training content and social clips.
- E-learning: narrated lessons in many languages, with word-by-word highlighting.
- Accessibility: spoken versions of documents, signs, kiosks and apps.
- Voice assistants and chatbots: speaking the replies of a language model in real time.
- Announcements and alerts: public address systems, IoT devices, alarms and mass notifications.
- Games and apps: character dialogue, with speech marks to drive lip sync.
Polly is HIPAA eligible under the AWS Business Associate Addendum, so healthcare apps can use it for patient-facing audio. The text you send may be stored and used only to provide and maintain the service, and you can opt out of its use for service improvement with an AWS Organizations AI services opt-out policy.
Jake: "So it reads anything you type, in any voice you pick, and nobody hears the van."
Ethan: "Any of about a hundred voices. And when the opening hours change, I change one line of text instead of booking you a quiet room."
The four voice engines: Standard, Neural, Generative and Long-form
Every Polly voice belongs to one or more engines, and the engine sets both the quality and the price. The same voice name, Joanna for example, can exist in several engines and sounds noticeably different in each.
| Standard | Neural | Generative | Long-form | |
|---|---|---|---|---|
| Price per million characters | $4 | $16 | $30 | $100 |
| How it sounds | Clear but robotic | Natural, smooth | Most human-like, expressive, conversational | Narrator style for long reading |
| Voices | About 60, the widest language list | More than 60 | 43 | 6 (4 US English, 2 Spain Spanish) |
| Speech marks | Yes | Yes | No | Yes |
| SSML | Every tag except newscaster | Most tags; newscaster style on some voices | Most tags, some partly | Most tags |
| Requests per second (default) | 80 | 8 | 8 | 8 |
| Best for | High-volume prompts, IoT, budgets | Most apps, e-learning, IVR | Narration, ads, assistants, videos | Audiobooks, articles, podcasts |
- Standard voices use the oldest technology (concatenative synthesis). They sound clearly synthetic, but they are cheap, fast, available in every Region and in the most languages, and support every SSML effect, including whispering and soft speech.
- Neural voices (NTTS) sound far more natural and are the default choice for most apps. A few, such as Joanna, Matthew, Amy and Lupe, also have a newscaster speaking style.
- Generative voices come from Polly's largest model, a billion-parameter transformer, and sound the most like a real person, with natural emphasis, pauses and emotion. Two things to know: they are available in 10 Regions (including US East, Oregon, Frankfurt, London, Tokyo and Sydney), and because the model is updated over time, the same text can sound slightly different months later, which matters for something like a podcast season recorded over a year.
- Long-form voices (Danielle, Gregory, Ruth and Patrick in US English; Alba and RaΓΊl in Spanish) are tuned to stay engaging over long passages, with narrator-like pacing and breathing, at a premium price.
How to choose: start with Neural. If the voice is the product (a video, an ad, a premium assistant), compare the same paragraph in Generative, which costs under twice as much. Use Long-form only when you have tried Generative on a long chapter and still prefer it. Use Standard for millions of short, functional prompts, or a language with no neural voice.
Amazon Polly voices: the English list
Polly has more than 100 voices across 42 languages and variants, including Arabic, Chinese (Mandarin and Cantonese), Dutch, French, German, Hindi, Italian, Japanese, Korean, Polish, Portuguese, Spanish, Swedish, Turkish and Welsh. Some are bilingual, such as Aditi and Kajal, which switch between Indian English and Hindi in the same sentence. These are the English voices and their engines:
| English variant | Voices (engines) |
|---|---|
| US English (en-US) | Danielle (generative, long-form, neural), Ruth (generative, long-form, neural), Matthew (generative, neural, newscaster), Joanna (generative, neural, standard, newscaster), Salli (generative, neural, standard), Stephen (generative, neural), Tiffany (generative only), Gregory (long-form, neural), Patrick (long-form only), Kendra, Kimberly and Joey (neural, standard), Ivy and Kevin (child voices, neural and standard), Justin (child voice, neural) |
| British English (en-GB) | Amy (generative, neural, standard, newscaster), Brian (generative, neural, standard), Emma (neural, standard), Arthur (neural) |
| Australian English (en-AU) | Olivia (generative, neural), Nicole and Russell (standard) |
| Indian English (en-IN) | Kajal (generative, neural, bilingual with Hindi), Aditi (standard, bilingual), Raveena (standard) |
| Irish, New Zealand, Singaporean, South African | Niamh, Aria, Jasmine and Ayanda (generative, neural) |
| Welsh English (en-GB-WLS) | Geraint (standard) |
Popular picks: Ruth and Matthew (generative) for narration, Danielle or Gregory for long reading, Joanna (neural) as an all-round voice, Amy and Brian for British English, Olivia for Australian English and Kajal for Indian English or Hindi. The DescribeVoices API returns the current list for any Region and engine, which is useful because voices are added from time to time.
Beyond the catalog, Brand Voice is a custom engagement in which AWS builds an exclusive neural voice for your company; it is arranged through your AWS account team rather than self-service.
Update, October 24, 2026: a new Generative model, 14 new voices, and Hindi and Finnish. On October 20, AWS put a new model behind Polly's Generative engine. It is a drop-in upgrade: the engine name, the voices you already use and your SSML stay the same, and numbers, email addresses and street addresses are meant to come out more reliably. The new voices include Kenneth and Nate (US English), Russell and Curtis (Australian English, so Russell in the list above now has a generative voice too), Kiara and Arjun (Hindi and Indian English), Tina (German), Carlos (US Spanish), Leo and Carolina (Brazilian Portuguese), Hyunduk (Korean) and Suvi (Finnish). Hindi and Finnish are on the Generative engine for the first time. Generative voices now run in 11 Regions: US East (N. Virginia), US West (Oregon), Frankfurt, Zurich, London, Sydney, Seoul, Singapore, Tokyo, Canada (Central) and AWS GovCloud (US-West). One thing to watch: because the model changed, a generative voice can sound slightly different from audio you made before October 20, so if a series must sound identical, keep the stored files or regenerate the whole set. To list what your Region offers, run aws polly describe-voices --engine generative.
Amazon Polly pricing and free tier
| Engine | Price per 1 million characters | Per 1,000 characters | Classic Free Tier allowance |
|---|---|---|---|
| Standard | $4.00 | $0.004 | 5 million characters a month |
| Neural | $16.00 | $0.016 | 1 million a month for 12 months |
| Generative | $30.00 | $0.03 | 100,000 a month for 12 months |
| Long-form | $100.00 | $0.10 | 500,000 a month for 12 months |
- What counts: every character of text you send for speech, including spaces and punctuation. SSML tags are not billed, so
<break time="1s"/>is free; only the text inside counts. - Speech marks (timing data for words and sentences) are billed at the same per-character rate as speech for standard, neural and long-form voices; generative voices do not offer them.
- No minimums, no subscriptions: you pay per request, billed monthly, in the same rates for real-time and background (asynchronous) synthesis.
- Free Tier: the allowances above apply to accounts on the classic AWS Free Tier. Accounts created from July 15, 2025 get up to $200 of Free Tier credits instead, which Polly can use.
- Other Regions: standard and neural prices are the same in other commercial Regions such as Europe (London); AWS GovCloud (US) charges $4.80 for standard and $19.20 for neural.
- Not part of Amazon Prime: Polly is an AWS service billed to an AWS account, unrelated to a Prime shopping membership.
- Replays are free: once you have the audio file, you can store it in S3 and play it a million times without paying Polly again.
A rule of thumb for converting: 1 million characters is about 23 hours of speech, so roughly 720 characters make a minute of audio, and a minute of narration costs about $0.003 standard, $0.012 neural, $0.022 generative or $0.072 long-form.
Worked examples: what Polly really costs
Example 1: the store's phone menu. Twenty prompts averaging 200 characters is 4,000 characters: 6 cents with a neural voice, and inside the free tier. When the hours change, regenerating one prompt costs a fraction of a cent.
Example 2: a news article. About 6,500 characters: $0.03 standard, $0.10 neural, $0.20 generative, $0.65 long-form.
Example 3: weekly video narration. A 10-minute script is roughly 7,200 characters: $0.22 per video with a generative voice. Sixteen videos a month is about 115,000 characters, about $3.46 a month generative, or $1.84 neural.
Example 4: an audiobook. An 80,000-word novel is around 480,000 characters, about 11 hours of audio: $48 long-form, $14.40 generative, $7.68 neural or $1.92 standard. Compare that with a human narrator, who typically charges per finished hour of audio.
Example 5: "listen to this article" on a blog, done wrong and done right. A blog with 200 posts of about 40,000 characters each is 8 million characters. Synthesizing each post once with a neural voice and storing the MP3 in S3 costs $128, one time, plus a few cents a month for storage. Calling Polly every time a visitor presses play, with 20,000 plays a month, would send 800 million characters a month: $12,800 a month for the same audio.
Ethan: "Same audio, same voice. One costs $128 once. The other costs $12,800 every month."
Jake: "Because the second one keeps asking Polly to read the same article again?"
Ethan: "Every single time someone presses play. Generate once, save the file, play the file."
Example 6: a voice assistant. A support assistant that speaks 10 replies of about 150 characters in each of 2,000 daily conversations sends 3 million characters a day, 90 million a month: $2,700 a month generative, $1,440 neural or $360 standard. Caching the most common replies (greetings, "please hold", opening hours) often cuts a third or more of that.
Example 7: an e-learning course. Fifty lessons of 5,000 characters each is 250,000 characters: $4 neural, and translating the course into four more languages with the same voices adds $16.
Try Amazon Polly in the console (the free demo)
- Sign in to the AWS console and open Amazon Polly, then choose Text-to-Speech (Polly's demo page).
- Choose the engine: Generative, Long-form, Neural or Standard.
- Choose the language and voice, for example English, US and Ruth.
- Type or paste your text, or switch on SSML to add pauses and pronunciation.
- Press Listen to hear it instantly.
- Choose Download to save an MP3 (or change the file format and sample rate in the settings), or Save to S3 for long text, which starts a background synthesis task.
- Open Lexicons to upload pronunciation rules, and apply them in the settings before listening again.
Console requests are billed like API requests, so listening repeatedly to a long text uses your free tier (or pennies). If your Region does not show the generative engine, switch to US East (N. Virginia) or another Region that has it.
Amazon Polly login, IAM permissions and access keys
There is no separate Polly account or login. Polly is part of your AWS account: sign in to the AWS Management Console (ideally as an IAM Identity Center or IAM user rather than the root email), type "Polly" in the search bar, and you are on the Text-to-Speech page. If you do not have an AWS account yet, creating one takes a few minutes and needs a payment card, even if you stay inside the free tier.
For code, your application needs AWS credentials with Polly permissions:
- On AWS compute (Lambda, EC2, ECS), attach an IAM role to the function or instance. No keys to store or rotate.
- On your own computer, run
aws configure ssoto sign in through IAM Identity Center, or create an IAM user with access keys (IAM, Users, your user, Security credentials, Create access key) and save them withaws configure. Keep keys out of your code and repositories. - In a browser or mobile app, get temporary credentials from an Amazon Cognito identity pool, or have your server create a short-lived presigned Polly URL, so no long-term key ever reaches the client.
A least-privilege policy for an app that only speaks text looks like this; the AWS managed AmazonPollyFullAccess policy is fine for experiments but grants far more:
{
"Version": "2012-10-17",
"Statement": [{
"Effect": "Allow",
"Action": ["polly:SynthesizeSpeech", "polly:DescribeVoices"],
"Resource": "*"
}]
}
Amazon Polly with Python (boto3) and the AWS CLI
The basic call returns an audio stream you write to a file. Your IAM user or role needs polly:SynthesizeSpeech:
import boto3
polly = boto3.client("polly", region_name="us-east-1")
resp = polly.synthesize_speech(
Engine="generative", # "standard", "neural", "long-form" or "generative"
VoiceId="Ruth",
OutputFormat="mp3", # "mp3", "ogg_vorbis", "pcm" or "json" (speech marks)
Text="Thanks for calling the furniture store. Press 1 for deliveries, or 2 for returns.",
)
with open("greeting.mp3", "wb") as f:
f.write(resp["AudioStream"].read())
Long text (more than 3,000 characters): start a background task that writes the audio straight to S3, up to 100,000 billed characters per task. It also needs s3:PutObject on the bucket, and can notify an SNS topic when it finishes:
with open("sofa-care-guide.txt", encoding="utf-8") as f:
guide = f.read()
task = polly.start_speech_synthesis_task(
Engine="long-form", VoiceId="Danielle", OutputFormat="mp3",
OutputS3BucketName="store-audio", OutputS3KeyPrefix="guides/sofa-care-",
Text=guide,
)
task_id = task["SynthesisTask"]["TaskId"]
# Later: check the status and get the S3 URL of the finished file
status = polly.get_speech_synthesis_task(TaskId=task_id)["SynthesisTask"]
print(status["TaskStatus"], status.get("OutputUri"))
From the command line, the output file name goes last:
aws polly synthesize-speech --engine neural --voice-id Joanna --output-format mp3 \
--text "Your sofa is out for delivery today." delivery.mp3
# List the generative voices for US English in this Region
aws polly describe-voices --engine generative --language-code en-US
To process many short texts, call synthesize_speech from a small pool of threads (no more than the engine's requests-per-second limit) or from a Lambda function triggered by new text files, and let the SDK's retries absorb occasional throttling. In boto3 you can make those retries smarter by creating the client with config=botocore.config.Config(retries={"mode": "adaptive", "max_attempts": 10}), which slows your request rate automatically when Polly starts throttling.
Amazon Polly in JavaScript (Node.js, AWS SDK v3):
import { PollyClient, SynthesizeSpeechCommand } from "@aws-sdk/client-polly";
import { writeFile } from "node:fs/promises";
const polly = new PollyClient({ region: "us-east-1" });
const out = await polly.send(new SynthesizeSpeechCommand({
Engine: "neural", VoiceId: "Joanna", OutputFormat: "mp3",
Text: "Your order is ready for pickup at the side entrance.",
}));
await writeFile("ready.mp3", Buffer.from(await out.AudioStream.transformToByteArray()));
In the browser, the @aws-sdk/polly-request-presigner package's getSynthesizeSpeechUrl function builds a signed, short-lived URL you can put straight into an <audio> element, using temporary credentials from a Cognito identity pool. For anything public, though, pre-generated MP3 files are cheaper and simpler than synthesizing in each visitor's browser.
Splitting long text yourself: if you prefer real-time calls over background tasks, split the text at sentence ends into pieces under 3,000 characters, synthesize each with the same voice and engine, and join the MP3 files in order. Joining MP3 files byte by byte usually plays fine; for a clean single file, re-encode the joined audio with a tool such as FFmpeg.
SSML: pauses, pronunciation, speed and emphasis
SSML (Speech Synthesis Markup Language) is a small set of XML tags that control how text is read. Send it with TextType="ssml" and wrap everything in <speak>:
<speak>
Thanks for calling the store. <break time="600ms"/>
Your order number is <say-as interpret-as="characters">AB12</say-as>,
arriving on <say-as interpret-as="date" format="mdy">10/24/2026</say-as>.
<prosody rate="90%">Press 1 for deliveries, or 2 for returns.</prosody>
Our <sub alias="Aal-borg">Aalborg</sub> sofas are in aisle four.
</speak>
<break>adds a pause, either a fixed time (time="750ms"ortime="2s", up to 10 seconds) or a strength fromx-weaktox-strong;<p>and<s>add paragraph and sentence pauses. Without SSML, plain punctuation already adds natural pauses: a comma gives a short one, a full stop a longer one, and a new paragraph longer still.<say-as>controls how dates, numbers, phone numbers, spelled-out characters and units are read.<sub>replaces a word with how it should be said;<phoneme>gives an exact IPA pronunciation.<prosody>changes speed and volume (and pitch on standard voices); it is only partly supported on neural, generative and long-form voices.<lang>reads a phrase with another language's pronunciation, such as a French dish name in an English sentence.<mark>places a named marker you get back in speech marks, for syncing visuals.<amazon:domain name="news">gives selected neural voices a newscaster style.- Standard voices only:
<emphasis>, whispering (<amazon:effect name="whispered">), soft phonation, breathing sounds and timbre changes.
SSML must be valid XML, so write & for "&" and < for "<" inside the text. A tag an engine does not support returns an error on standard, neural and long-form voices, which is the most common cause of a sudden InvalidSsmlException after switching engines.
Pronunciation lexicons for brand names and jargon
When the same word is mispronounced everywhere (a product name, a surname, a medical term, an acronym), fix it once with a lexicon instead of wrapping every occurrence in SSML. A lexicon is a small XML file in the W3C PLS format:
<?xml version="1.0" encoding="UTF-8"?>
<lexicon version="1.0" xmlns="http://www.w3.org/2005/01/pronunciation-lexicon"
alphabet="ipa" xml:lang="en-US">
<lexeme><grapheme>Aalborg</grapheme><alias>All-borg</alias></lexeme>
<lexeme><grapheme>SKU</grapheme><alias>skew</alias></lexeme>
</lexicon>
aws polly put-lexicon --name storewords --content file://storewords.pls
# then pass LexiconNames=["storewords"] in synthesize_speech
Each entry maps a written word (the grapheme) to either an alias, another spelling that Polly already says correctly, which is the easy option, or a phoneme, an exact pronunciation written in IPA or X-SAMPA symbols, which is precise but harder to write. Aliases also expand abbreviations, such as "W3C" to "World Wide Web Consortium". Matching is case-sensitive in some languages, so add both "Aalborg" and "AALBORG" if your text uses both.
Lexicon or SSML? Use a lexicon for words that appear across many texts (brand names, staff names, product codes), because it applies automatically to every request that names it. Use SSML <sub> or <phoneme> for a one-off word in a single script. You can manage lexicons with list-lexicons, get-lexicon and delete-lexicon, or on the console's Lexicons page, where you can also test a lexicon before using it.
You can store up to 100 lexicons per account per Region, each up to 40,000 characters, and apply up to five to one request; very large lexicons slightly increase the time to the first audio. Lexicons are Regional, so upload them in every Region where you synthesize speech.
Speech marks: word highlighting and lip sync
Speech marks are timing data instead of audio: for each word, sentence, SSML mark or mouth shape (viseme), Polly tells you when it is spoken, in milliseconds. Apps use them to highlight words as they are read (karaoke style, great for e-learning and accessibility) and games use visemes to move a character's lips. Request them with the same text and voice as the audio, with OutputFormat="json":
marks = polly.synthesize_speech(
Engine="neural", VoiceId="Joanna", OutputFormat="json",
SpeechMarkTypes=["sentence", "word"], Text="Our sofas ship in five days.",
)
for line in marks["AudioStream"].read().decode().splitlines():
print(line) # {"time":6,"type":"word","start":0,"end":3,"value":"Our"} ...
There are four types: sentence and word marks (with the time and the character offsets of each in your text), viseme marks (the mouth shape at each moment, for animation), and ssml marks (the names of your <mark> tags, for triggering slides or animations at exact points). To highlight words as audio plays, load the word marks next to the MP3, and on each tick of the audio player compare its current time in milliseconds with the marks to find the word being spoken, then highlight the text between that mark's start and end offsets.
Speech marks are billed per character like speech, so getting both audio and marks for a text costs twice. Generative voices do not support them; use a neural voice when you need highlighting. Make both requests with exactly the same text, voice, engine and SSML, or the timings will not match the audio.
Real-time speech and bidirectional streaming for voice assistants
SynthesizeSpeech already streams: audio starts arriving within a fraction of a second, so you can begin playback before the whole response is generated. That is enough for phone menus and apps that speak complete sentences.
For assistants built on a large language model, the reply itself arrives word by word, and waiting for the full reply before speaking adds seconds of silence. Polly's bidirectional streaming API, StartSpeechSynthesisStream, solves this for generative voices: you open one connection, send text as the model produces it, and receive audio on the same connection as soon as there is enough text to speak, with a flush event to force synthesis of buffered text and a close event when the reply ends. It runs over HTTP/2 event streams, allows up to 8 new streams per second and 8 at once by default, and is available in US East (N. Virginia), US West (Oregon), Frankfurt, London, Zurich, Singapore, Sydney and Canada (Central). Check that your AWS SDK version supports it before planning around it.
Three things keep a voice assistant feeling quick: send text in sentence-sized pieces rather than single words, start playback as soon as the first audio chunk arrives, and cache fixed phrases so they play with no synthesis at all.
Add a "listen to this article" button to your website
- Generate audio when you publish, not when someone listens: a Lambda function triggered by your CMS (or by a new text file in S3) strips the HTML, splits the text into chunks under 3,000 characters or starts a synthesis task, and saves one MP3 per post.
- Store the MP3 in S3, named after the post's URL, in a bucket that only your CDN can read; our guide on giving access to an S3 bucket covers the bucket policy.
- Serve it through CloudFront so playback is fast everywhere and costs a fraction of a cent per listen.
- Add a player to the post template:
<audio controls preload="none" src="https://cdn.example.com/audio/my-post.mp3"></audio>, withpreload="none"so visitors who never press play download nothing. - Regenerate on edits: when a post changes, the same function replaces the MP3, costing only that post's characters.
A 6,500-word article is about 40,000 characters: $0.64 neural or $1.20 generative, once. If your site runs on a hosted platform with no server-side code, generate the MP3 on your own computer with the CLI or the console and upload it with the post.
Amazon Polly vs other text-to-speech options
- Windows and Office built-in voices: free and offline, good for listening to your own documents (see our guides to changing the Windows voice and Read Aloud in Word), but they are tied to one PC and are not an API you can call from a server.
- Google Cloud Text-to-Speech and Azure AI Speech: the closest equivalents, also billed per character with standard, neural and premium voice tiers and custom voice options. Pick the one in the cloud you already use; quality is close at the top tiers, so listen to your own text in each.
- ElevenLabs and similar voice platforms: very expressive voices and easy voice cloning, sold mainly as monthly plans with character allowances. Strong for creators and characters; at very high volume, per-character cloud prices are usually lower.
- Model providers' speech APIs: several AI labs now sell text-to-speech models alongside their language models, such as Microsoft's MAI-Voice-2.1 (covered in our MAI-Transcribe-2 guide) and Mistral's Voxtral family.
- Open-source models you run yourself: free per character and private, running on your own PC or server; you handle hosting, scaling and quality checks.
Where Polly wins: it lives inside AWS with IAM permissions, CloudWatch metrics, HIPAA eligibility and the same bill as the rest of your stack; prices are simple and per character with no plan to outgrow; it has four quality tiers so you can mix cheap prompts with premium narration; and SSML, lexicons, speech marks and streaming are all built in. Where others may win: voice cloning from a short sample, the very widest choice of character voices, or zero cost for offline personal use.
Polly vs Transcribe vs Lex, and Polly in Amazon Connect
AWS has several speech services, and their names do not explain themselves:
- Amazon Polly speaks: text in, audio out.
- Amazon Transcribe listens: audio in, text out, for call recordings, subtitles, meeting notes and live captions. If you searched "Polly speech to text", Transcribe is the service you want; our guides to offline speech to text and Mistral's Voxtral models cover alternatives.
- Amazon Lex converses: it builds chatbots and voice bots that understand what a person wants (an intent such as "track my order"), ask follow-up questions and call your code. Voice bots built with Lex listen with speech recognition and reply with Polly voices.
- Amazon Connect is the contact center that ties them together: callers hear Polly prompts, Lex bots answer routine questions, and Transcribe-based analytics write up the calls.
Polly in Amazon Connect: any prompt in a contact flow can be typed text instead of a recording. The Set voice block picks the language, voice and speaking style, and the Play prompt and Get customer input blocks accept plain text or SSML, so you can add pauses or spell out order numbers. Changing a message is a text edit and a publish, with no recording session.
Jake: "So Polly is the mouth, Transcribe is the ears, and Lex is the part that figures out what people mean."
Ethan: "And Connect is the building they all work in."
Is Amazon Polly good? An honest verdict
For most business uses, yes. The neural and generative voices are good enough that many listeners do not notice they are synthetic in a phone menu or a training video, the price is low and predictable, and nothing else in AWS makes adding speech to an app easier. Polly's weaker points: the voice catalog is smaller than some specialist platforms, there is no self-service voice cloning, generative voices lack speech marks and are in fewer Regions, and the standard voices now sound dated. Dramatic or character-heavy performances (audio drama, game characters with strong emotions) can still sound better on platforms built for that. If you are already on AWS, start with Polly's neural and generative voices and only move if your own listening tests say so.
Does Polly use AI? Yes. Neural voices use deep neural networks, and the generative engine is generative AI in the full sense: a billion-parameter transformer model that produces speech codes, then a decoder that turns them into audio. Only the standard voices use the older, pre-AI method of stitching recorded sounds together.
Can I use Amazon Polly audio commercially, on YouTube or in apps?
Yes. As between you and AWS, the audio Polly generates belongs to you, and you can store, cache and replay it with no extra fee and no limit on how often it is played. That covers videos, podcasts, apps, games, phone systems, e-learning and advertising. Two conditions still apply: you need the rights to the text you convert (a book you did not write, for example, needs the rights holder's permission for an audiobook), and the platforms you publish on have their own rules, such as YouTube's monetization policies on repetitive or mass-produced content and its disclosure rules for realistic synthetic media, which apply whatever voice you use.
The store's assembly videos went on YouTube with a generative voice, and the narration cost less than the coffee Jake drank while waiting to be asked to record them.
Limits and quotas
- Text per request: up to 3,000 billed characters (6,000 including SSML tags) for
SynthesizeSpeech; up to 100,000 billed characters (200,000 total) for aStartSpeechSynthesisTaskjob. - Audio length:
SynthesizeSpeechoutput stops at 10 minutes; use a synthesis task for longer audio. - Requests per second: standard voices 80 (burst 100); neural and long-form 8 (burst 10); generative 8. Background tasks: 10 per second for standard and neural, 1 for generative and long-form. Several are adjustable through Service Quotas.
- Concurrent requests: 80 standard, 18 neural, 26 generative, 26 long-form.
- Lexicons: 100 per account, 40,000 characters each, five per request.
- SSML: breaks of up to 10 seconds each, speaking rate no lower than -80 percent, and no support for the
audio,lexicon,lookuporvoicetags. - Formats: MP3 and OGG Vorbis at sample rates of 8, 16, 22 or 24 kHz (24 kHz is the default for generative voices), raw PCM at 8 or 16 kHz, plus JSON for speech marks.
Common Amazon Polly errors and fixes
- ThrottlingException (HTTP 400): you exceeded the engine's requests per second, most often with neural or generative voices at 8 per second. Retry with backoff and jitter, limit the number of parallel workers, cache repeated phrases, or request a quota increase.
- TextLengthExceededException: more than 3,000 billed characters in one
SynthesizeSpeechcall. Split the text at sentence boundaries and join the audio files, or useStartSpeechSynthesisTask. - InvalidSsmlException: the SSML is not valid XML (an unescaped & or <, a missing
<speak>),TextTypeis not set tossml, or the tag is not supported by the engine, such as<emphasis>or whispering on a neural voice. - EngineNotSupportedException or "voice does not support the selected engine": the voice does not exist in that engine (Tiffany is generative only, Patrick long-form only), or the Region has no generative or long-form voices. Check with
describe-voicesin your Region. - MarksNotSupportedForFormatException: speech marks need
OutputFormat="json", and generative voices do not produce them. - LexiconNotFoundException: the lexicon name is misspelled or was uploaded to another Region.
- AccessDeniedException: the role lacks
polly:SynthesizeSpeech, or, for synthesis tasks,polly:StartSpeechSynthesisTaskpluss3:PutObjecton the output bucket (andsns:Publishif you set a topic). - Audio stops after exactly 10 minutes: the real-time output limit; switch to a synthesis task.
- A generative voice cuts off a word at the end of a sentence: a built-in safety stop occasionally ends speech early; regenerate that sentence, or rephrase it slightly.
- A word is pronounced wrong every time: add it to a lexicon, or use
<sub>or<phoneme>in SSML.
Best practices and ways to cut the bill
- Generate once, play many times: store audio in S3 and serve it from a CDN; replays cost nothing in Polly.
- Cache fixed phrases in voice apps and assistants, and synthesize only the dynamic parts.
- Match the engine to the job: standard for bulk functional prompts, neural for most content, generative where the voice matters most, long-form for long reading.
- Keep voices consistent across a series by storing the voice ID, engine and SSML with each file, and remember generative voices can drift slightly as the model is updated.
- Use lexicons for names and jargon so pronunciation is consistent everywhere.
- Retry throttled requests with backoff and keep parallel calls under the per-engine limit.
- Watch usage in CloudWatch (
RequestCharacters,ResponseLatency, error counts) and set an AWS billing alert before you launch anything that synthesizes per user. - Grant least privilege: apps usually need only
polly:SynthesizeSpeech.
Frequently asked questions
What is Amazon Polly?
AWS's text-to-speech service. It converts text into natural-sounding speech in more than 100 voices across 42 languages and variants, returning MP3, OGG or PCM audio through an API or the console.
How much does Amazon Polly cost?
Per million characters: $4 for standard voices, $16 for neural, $30 for generative and $100 for long-form. One million characters is about 23 hours of speech, so a typical article costs between 3 cents and 65 cents.
Is Amazon Polly free?
Partly. Accounts on the classic Free Tier get 5 million standard characters a month, and for the first 12 months 1 million neural, 500,000 long-form and 100,000 generative characters a month. Accounts created from July 15, 2025 use Free Tier credits instead. Beyond that, you pay per character.
What is the difference between neural and generative voices in Polly?
Neural voices sound natural and cost $16 per million characters. Generative voices come from Polly's largest model and sound more human, with more natural emotion and emphasis, for $30 per million characters, but they do not support speech marks and are available in fewer Regions.
What are the best Amazon Polly voices?
For US English narration, the generative voices Ruth, Matthew, Danielle and Stephen; for long reading, the long-form Danielle and Gregory; for British English, Amy and Brian; for Australian English, Olivia; for Indian English and Hindi, Kajal. Listen to your own text in two or three before deciding.
How many languages does Amazon Polly support?
42 languages and language variants, from Arabic and Cantonese to Welsh, with several English, Spanish, French, German and Portuguese variants. Not every language has neural or generative voices.
Can I use Amazon Polly voices commercially?
Yes. The audio you generate belongs to you, and you can store and replay it as often as you like at no extra cost, including in videos, apps, games and phone systems. You need the rights to the text, and the platforms you publish on have their own rules.
Can I use Amazon Polly for YouTube videos?
Yes. The narration is yours to use in videos. YouTube's own policies on monetization and on disclosing realistic synthetic content still apply, whatever voice you use.
What is the character limit in Amazon Polly?
3,000 billed characters (6,000 including SSML tags) per real-time request, and 100,000 billed characters per background synthesis task, which writes the audio to S3. Real-time audio is also capped at 10 minutes.
Are SSML tags billed in Amazon Polly?
No. Only the text characters are billed; SSML tags such as breaks, prosody and say-as are free, so adding pauses and pronunciation hints does not raise the cost.
How do I use Amazon Polly in Python?
Create a boto3 Polly client, call synthesize_speech with an Engine, VoiceId, OutputFormat and Text, and write the AudioStream to a file. For long text, call start_speech_synthesis_task with an S3 bucket for the output.
Does Amazon Polly support Hindi and Indian English?
Yes. Kajal is a bilingual generative and neural voice for Indian English and Hindi, and Aditi is a bilingual standard voice. Raveena is a standard Indian English voice.
Can Amazon Polly clone my voice?
Not as a self-service feature. Polly offers a fixed catalog of voices, plus Brand Voice, a custom engagement in which AWS builds an exclusive voice for a company, arranged through your AWS account team.
What output formats does Amazon Polly support?
MP3 and OGG Vorbis audio at 8, 16, 22 or 24 kHz, raw PCM at 8 or 16 kHz (handy for telephony), plus JSON speech marks with timing for words, sentences, visemes and SSML marks.
What are Amazon Polly speech marks?
Timing data that tells you when each word, sentence or mouth shape occurs in the audio, in milliseconds, used for word highlighting and lip sync. They are billed like speech and are available for standard, neural and long-form voices.
Why is my Polly request throttled?
Neural, long-form and generative voices allow about 8 requests per second by default, and standard voices 80. Retry with backoff, cap your parallel requests, cache repeated phrases, or request a higher quota.
Can Amazon Polly read long documents or books?
Yes, with StartSpeechSynthesisTask, which handles up to 100,000 characters per task and saves the audio to S3. For a whole book, split it into chapters and run one task per chapter.
Is Amazon Polly HIPAA eligible?
Yes. Polly is a HIPAA eligible service under the AWS Business Associate Addendum, so it can be used for patient-facing audio when your account has a BAA in place.
Is Amazon Polly free with Prime?
No. Polly is an Amazon Web Services product billed to an AWS account, separate from an Amazon Prime shopping membership. Its free allowances come from the AWS Free Tier instead.
How do I log in to Amazon Polly?
Sign in to the AWS Management Console with your AWS account and search for Polly. There is no separate Polly login; for code, use an IAM role or access keys with polly:SynthesizeSpeech permission.
How do I add a pause in Amazon Polly?
Use SSML with a break tag, such as <break time="1s"/> inside <speak>, up to 10 seconds per break, and set the text type to SSML. Commas, full stops and new paragraphs also add natural pauses without SSML.
Is Amazon Polly down?
Check the AWS Health Dashboard for your Region. Polly runs separately in each Region, so an issue in one Region rarely affects others; if only your requests fail, check throttling, permissions and the Region your code uses.
Does Amazon Polly work in real time for chatbots?
Yes. SynthesizeSpeech streams audio as it is generated, and the bidirectional streaming API for generative voices lets you send text while your chatbot is still writing its reply and receive audio on the same connection.
For readers preparing for an AWS certification
We have been covering AWS certifications alongside our fixes, so here is a short extra for readers studying for one. If you only came for the fix, you already have it; feel free to skip this part.
Exams that test this topic: AI Practitioner (AIF-C01) • Machine Learning Engineer – Associate (MLA-C02) • Cloud Practitioner.
Here are the exam-style questions this topic tends to raise, with the reasoning you need to answer them:
1. An application must convert product descriptions into natural-sounding spoken audio. Which AWS service fits?
Amazon Polly (text-to-speech).
2. Call recordings must be converted into text transcripts. Which service fits, rather than Polly?
Amazon Transcribe (speech-to-text).
3. A company name is mispronounced in every Polly request across many scripts. What is the most efficient fix?
A pronunciation lexicon applied to the requests (or SSML sub/phoneme for one-off words).
4. An e-learning app must highlight each word as Polly reads it. Which feature provides the timing?
Speech marks (word and sentence marks, OutputFormat json) from a neural or standard voice.
5. A 40,000-character article must be converted to one audio file without splitting it in code. Which API fits?
StartSpeechSynthesisTask, which writes the audio to S3 (up to 100,000 billed characters).
6. A website plays the same Polly narration to thousands of visitors. How should costs be minimized?
Generate the audio once, store it in S3 and serve it through CloudFront; replays cost nothing in Polly.
7. A conversational bot must understand intents and ask follow-up questions by voice. Which service provides the conversation, and which the voice?
Amazon Lex for the conversation, Amazon Polly for the spoken replies.
Exam codes and dates change often. Our AWS certification changes page keeps track of what is current.
The new phone menu went live the same afternoon, read by Ruth, with a 600-millisecond pause before the options and the store's Aalborg sofas finally pronounced the way the owner said them. The sofa care guide became a seven-minute MP3 on the website, and the three assembly videos got their narration for less than a dollar. Total Polly bill for the month: $0.00, all inside the free tier. Jake asked whether Ruth could also take the delivery calls. "She can read the answers," Ethan said. "Someone still has to drive the van."
📌 If you keep one line from this page
Amazon Polly turns text into speech for $4 (standard), $16 (neural), $30 (generative) or $100 (long-form) per million characters, about 23 hours of audio; the audio is yours, so generate it once, store it in S3 and replay it free, and use SSML and lexicons to get pauses and pronunciation right.
Neural for most jobs · generative where the voice matters · cache everything.
Revision note. Written October 21, 2026. Prices are US East (N. Virginia) list prices on that date. We will update this page when Polly prices, engines or voices change.