AI Voice API Pricing Compared

Disclosure: This post contains affiliate links; we may earn a commission at no extra cost to you.

AI Voice API Pricing Compared: ElevenLabs, Amazon Polly, Google Cloud, and Azure

Amazon Polly is the cost winner for high-volume conventional text-to-speech: Standard costs $4 per million characters, Neural $16, Generative $30, and Long-Form $100 under current US list pricing. ElevenLabs is much more expensive per character but earns its premium with expressive voices, strong cloning, low-latency models, and creator-friendly controls. Google Cloud Text-to-Speech closely matches Polly’s $4 Standard and $16 WaveNet tiers, while Microsoft Azure is compelling for organizations that need multilingual neural voices, SSML styles, and Microsoft cloud governance.

Cost alone does not determine the right API. A phone assistant needs latency, streaming, μ-law audio, interruption handling, and stable concurrency. An audiobook needs long-form consistency and pronunciation tools. A navigation system may prefer a cheap deterministic voice. This comparison uses published list prices as a budgeting snapshot; region, contract, model, and free-tier rules can change, so verify the official calculator before deployment.

Best premium API: ElevenLabs

ElevenLabs’ 2026 pay-as-you-go API pricing lists Flash/Turbo text-to-speech at about $0.05 per 1,000 characters and Multilingual v2/v3 at about $0.10 per 1,000 characters. That equals roughly $50 or $100 per million—several times the cost of Polly Neural. The audio quality, expressive range, professional voice cloning, and latency can justify it for customer-facing speech.

Our pick: ElevenLabs API

ElevenLabs also sells subscription tiers with included characters at a better effective rate as volume rises. Current plans include Free, Starter around $6, Creator around $22, Pro around $99, Scale around $299, Business around $990, and Enterprise quotes. The API pricing calculator shows model-specific included characters, not one universal credit conversion. Run both subscription and PAYG scenarios.

Published API prices at a glance

Provider/model	Current public unit	Approximate cost per 1M characters	Best use
Amazon Polly Standard	$4 per 1M characters	$4	Announcements, accessibility, utility narration
Amazon Polly Neural	$16 per 1M	$16	Higher-quality general application speech
Amazon Polly Generative	$30 per 1M	$30	Expressive real-time voices in supported locales
Amazon Polly Long-Form	$100 per 1M	$100	Long articles and publication-oriented narration
Google Cloud Standard	$0.000004 per character after free tier	$4	Low-cost cloud TTS
Google Cloud WaveNet	$0.000016 per character after free tier	$16	Natural general-purpose speech
ElevenLabs Flash/Turbo PAYG	$0.05 per 1K characters	$50	Low-latency premium speech
ElevenLabs Multilingual v2/v3 PAYG	$0.10 per 1K characters	$100	Expressive and multilingual premium content
Microsoft Azure Speech	Per character, model and region	Check current regional calculator	Enterprise multilingual and SSML-controlled speech

These figures exclude tax, negotiated discounts, storage, egress, telephony, speech recognition, LLM tokens, orchestration, support, custom voice training, and engineering. Google counts spaces and most SSML tags. AWS charges Speech Marks characters as applicable. Model availability differs by region and language.

Translate characters into audio hours

AWS gives a useful benchmark: one million characters is approximately 23 hours and eight minutes for its pricing examples. Actual duration varies with language, speaking rate, punctuation, numbers, and SSML. A conservative English planning factor is 800–1,000 characters per minute.

Using the AWS example, a million-character bill produces roughly:

• Polly Standard: about $0.17 per raw audio hour.
• Polly Neural: about $0.69 per raw hour.
• Polly Generative: about $1.30 per raw hour.
• Polly Long-Form: about $4.32 per raw hour.
• ElevenLabs Flash PAYG: about $2.16 per raw hour.
• ElevenLabs Multilingual PAYG: about $4.32 per raw hour.

Do not use these as finished-production costs. Pronunciation retries, emotional alternatives, failed requests, dynamic prompts, and regenerated paragraphs increase usage. Add 10–20% for stable utility speech and 30–100% for expressive content. Caching reduces repeat synthesis: generate “Your payment was received” once rather than on every call if the license and application allow it.

ElevenLabs API: quality, cloning, and low latency

Eleven Flash v2.5 is designed for real-time speed, with official documentation quoting latency around 75 milliseconds in favorable conditions. It supports many languages, long request text, streaming, and telephony-friendly formats. Multilingual v2 favors stability and nuanced long-form quality. Eleven v3 provides richer expression and audio tags but has different request limits and production characteristics.

Output formats include MP3, PCM at multiple sample rates, Opus, μ-law, and A-law depending on tier and endpoint. This matters: Twilio and other telephony systems often want 8 kHz μ-law, while editing and games may prefer PCM. Avoid paying network and compute costs to transcode unnecessarily.

The API exposes request identifiers and character-cost metadata in response headers, which should be logged for cost attribution. Use per-feature budgets and preserve trace IDs for disputed consumption. Free regeneration may be available when the same text and settings meet current conditions; do not design a retry loop on an assumption that every repeat is free.

PAYG funds are prepaid, nonrefundable, and currently expire 12 months after purchase. A balance can coexist with a subscription; included credits are used first. Auto Top Up maintains service but can create surprise spend if there is no monthly cap. Set a cap, alerts, and a hard application quota. Free-tier PAYG still inherits plan-level voice-slot and concurrency limits, even though it unlocks paid API calls.

For production, higher subscription tiers buy more than characters: increased concurrency, priority, clone slots, seats, support, SSO, data terms, and enterprise options such as zero-retention mode. Enterprise zero retention disables history-dependent features. Test how that affects request stitching and debugging.

Amazon Polly: simplest high-volume economics

Polly charges only for characters processed and allows generated speech to be cached and replayed without additional synthesis charges. There is no required monthly subscription. Standard, Neural, Long-Form, and Generative engines have different voice catalogs, regions, quotas, and SSML support.

Current official prices are unusually clear: $4, $16, $100, and $30 per million respectively outside the applicable free tier. AWS has historically offered first-year monthly free character allowances per engine, while the broader AWS Free Tier changed for new accounts in 2025 to include selectable plans and credits. Check account creation date and current program; never assume an old “12 months free” tutorial applies.

Polly’s Generative engine expanded in 2026 with additional voices/locales, regions, and bidirectional streaming support. Bidirectional streaming helps an application feed LLM output while receiving audio rather than waiting for the entire text. Regional support still matters for data residency and latency.

Quotas vary by engine. Neural and Generative requests have lower transaction/concurrency limits than Standard, and an unexpected launch can throttle. Request increases before production, implement exponential backoff, and monitor errors. Do not split text into tiny requests without testing prosody; sentence fragmentation can make speech choppy.

Polly is ideal for huge catalogs, accessibility, IVR prompts, public-sector announcements, and embedded products that value predictable cost. It lacks the creator marketplace and expressive cloning workflow that makes ElevenLabs attractive. A custom brand voice requires a different service or enterprise path.

Google Cloud Text-to-Speech: comparable unit price, broad Google integration

Google Cloud bills by characters submitted each month. The official table lists Standard voices with four million free characters monthly and $4 per million beyond that, and WaveNet with one million free and $16 per million beyond that, subject to current program and SKU. Other voice families—such as Neural2, Studio, Chirp or newer models—may have distinct rates; inspect the current pricing table rather than applying WaveNet pricing to all premium voices.

Spaces count. Most SSML tags count, with documented exceptions such as `mark`. A verbose SSML wrapper repeated around every phrase can increase billed text. Google supports REST and gRPC, speaking-rate and pitch adjustment, volume gain, audio profiles, and formats including MP3, Linear16, and OGG Opus.

Google is attractive when the application already uses Cloud Run, Firebase, Dialogflow, Vertex AI, or Google identity/billing. Other resources are billed separately. Create a dedicated project, restrict the API key or service account, set quotas, and export billing data. A compromised unrestricted credential can consume the free tier in minutes.

Voice names and regional availability can change. Pin a voice/model in configuration and run regression audio before upgrades. Store an approved set of reference phrases for pronunciation and timbre.

Microsoft Azure Speech: strongest SSML and enterprise fit

Azure bills standard synthesis per character with prices varying by model, region, and customer agreement. The official page lists free monthly capacity for selected tiers and directs users to the Azure calculator for exact rates. HD voices, custom neural voices, avatar or real-time products, and batch speech may have separate pricing.

The service’s advantage is control. Supported neural voices accept SSML for rate, pitch, pronunciation, breaks, styles, roles, and style degree. A developer can create a calm customer-service response or newscast delivery without probabilistic bracket prompts. The language-support matrix identifies which exact voices implement which styles.

Azure fits organizations already using Entra ID, Azure Monitor, private networking, regional resources, and Microsoft procurement. Custom Neural Voice adds identity and data governance but is limited-access and quote-driven. Budget talent consent, studio recording, training, deployment endpoint, and operational review alongside synthesis.

Calculator estimates can omit support plans, networking, storage, Functions, or monitoring. Tag resources and create cost alerts. If speech is generated from a large language model, include Azure OpenAI or other model tokens separately.

Latency and streaming can outweigh unit price

Batch audiobook synthesis can wait seconds for high-quality output. A voice assistant cannot. Measure time to first byte, time to first playable audio, real-time factor, jitter, and failure rate from the user’s region. Test short acknowledgments, long sentences, numbers, and tool-call delays.

Streaming reduces perceived delay but creates pronunciation challenges when later context would change how a word is read. Feed complete clauses where possible. Pre-generate stable greetings and disclosure. Cache common responses at the edge, respecting data and license constraints.

Voice-agent cost includes speech-to-text, LLM, TTS, telephony, and platform orchestration. ElevenAgents lists audio-minute rates separately from model generation under relevant plans; AWS Connect/Lex/Bedrock or Azure Communication Services/OpenAI similarly add components. Price 100 real calls, not a single TTS request.

Reliability costs

Cloud APIs provide service quotas, but business continuity still requires design. Implement provider timeouts, retries with idempotency, circuit breakers, cached fallbacks, and an emergency stock voice. Avoid automatic cross-provider failover for a cloned executive voice unless consent and brand review cover the fallback.

Log sanitized input length, model, voice ID, region, latency, request ID, character cost, output duration, and error. Do not log sensitive text by default. Separate personally identifiable content from observability. Enterprise contracts may offer data residency, zero retention, BAAs, SLAs, support, and indemnity unavailable at self-service rates.

Cost scenario: one million characters per month

For roughly 23 raw audio hours, Polly Standard is about $4, Polly Neural $16, Polly Generative $30, Google Standard $4 after free use, and WaveNet $16 after free use. ElevenLabs PAYG is approximately $50 on Flash/Turbo or $100 on Multilingual v2/v3. Azure requires the region/model calculator.

At this low volume, engineering cost dwarfs synthesis. Choose voice quality and operational fit. A $84 difference is irrelevant if the less suitable voice lowers conversion or needs a week of custom integration.

Cost scenario: 100 million characters per month

At list price, Polly Neural is around $1,600 and Polly Generative around $3,000; ElevenLabs Flash PAYG around $5,000 and Multilingual around $10,000 before subscription or enterprise volume economics. Google WaveNet list pricing would be in Polly Neural’s range after free allowance, while other Google models differ.

At this scale, request enterprise quotes. Compare committed-use discounts, concurrency, region capacity, support, data terms, custom voices, and outage credits. A 10% price difference is meaningful, but so is a rate limit that blocks launch traffic or a voice change that requires reapproval in 20 languages.

How to run a fair bake-off

Use 100 representative sentences across every language: names, addresses, acronyms, dates, currency, emotional lines, long paragraphs, and edge cases. Generate at matched sample rate and loudness. Blind-rate intelligibility, naturalness, brand fit, consistency, and pronunciation. Measure latency and billable characters from actual responses.

Calculate cost after retries and cache hits. Add developer hours, infrastructure, monitoring, security, support, and review. Verify commercial rights, voice consent, data retention, and output ownership. Then run a week-long load test below and above expected peak.

Polly wins pure list-price predictability, Google offers nearly identical basic economics inside its ecosystem, Azure wins deterministic enterprise speech controls, and ElevenLabs earns its premium when natural expressive speech or verified cloning affects the product. The cheapest API is the one that meets quality and reliability requirements without forcing the team to rebuild the surrounding system.

Related reading: AI Voice Pricing Models Compared · AI Voice Emotion Controls Compared · Free AI Voice Tools That Sound Good