2026 TTS Pricing Comparison: 11 Providers Compared (Free & Paid)

A useful TTS pricing comparison cannot reduce every provider to one “cost per million characters.” In 2026, some products meter characters, others use shared credits, seconds, tokens, seats, or subscription features. Free allowances may apply only to selected models, new accounts, or personal listening.

This guide compares eleven options by billing structure first, then shows how to build a reproducible project estimate. Prices and plans change; verify the official checkout or pricing page before purchase.

Quick Comparison

Provider or route	Primary billing model	Free access	Important cost boundary
OfflineTTS	No service charge; local device resources	Browser tool has no paid plan	Manual operation, hardware, storage, and review time
ElevenLabs	Subscription with shared credits; product-dependent usage	Evaluation tier	Credits are not one universal character allowance
Google Cloud TTS	Model-specific characters or tokens	Model-specific allowances may apply	Voice families have different units and rates
Azure Speech	Metered service by feature and model	Azure free tier may apply	Region, voice class, and custom features matter
Amazon Polly	Characters by Standard, Neural, Long-Form, or Generative class	Account/credit rules vary	Each engine class has a separate rate
PlayHT	Subscription and/or API plan	Check current plan	Product and concurrency limits affect usable volume
Murf	Subscription workspace and usage allowance	Trial access may apply	Seats, projects, and voice-generation limits matter
WellSaid	Subscription/workspace plans	Check current offer	Seat and production features matter alongside minutes
Speechify	Consumer reading subscription; separate products may differ	Free product access	Reading, creator, and API use are not one license
NaturalReader	Personal plans plus separate Commercial generator	Free personal voices	Personal audio is not licensed for redistribution
Self-hosted Kokoro	Hardware or rented compute	Model is available under its model-card license	Engineering, uptime, storage, and operations are paid indirectly

The “Commercial?” column common in older tables hides too much. Commercial rights can depend on the exact plan, model asset, cloned voice consent, source text, and output use. Treat licensing as a separate review item.

Provider-by-Provider Breakdown

OfflineTTS — User-Operated Browser Generation

OfflineTTS does not charge for characters and has no subscription plan. Selected engines run in the browser after required assets load. This makes it easy to audition and export narration without an account, but it is not a hosted API or an uptime-backed service.

Budget for the user’s device, model downloads, browser storage, failed generations, editing, and review. A long manuscript may need to be split into smaller sections, and performance varies by engine and hardware. The engine comparison documents those workflow differences.

ElevenLabs — Shared Credits Across Products

ElevenLabs sells subscription tiers with credits shared across several Creative products. TTS, dubbing, transcription, music, and other features can consume that pool differently. Voice model choice may also change consumption. Converting the headline plan to “characters per dollar” without documenting the model and product can produce the wrong estimate.

Record the plan’s renewal price, included credits, the selected model’s usage rate, commercial terms, overage rules, and expected re-renders. Voice cloning and Studio can be the reason to subscribe even when raw TTS volume is not the cheapest line item.

Google Cloud TTS — Model-Specific Characters and Tokens

Google Cloud’s official page lists multiple voice families with different pricing. Some are character-based, while newer offerings may use token-based input and audio-output accounting. Historical Standard and WaveNet examples are not a complete representation of the current catalog.

Estimate the exact model from current documentation. Include punctuation and SSML according to Google’s billing definition, then add retries, storage, and application operations.

Azure Speech — Feature, Model, and Region Matter

Azure’s Speech pricing covers more than one synthesis option. The relevant rate can depend on neural voice class, region, deployment mode, custom features, and commitment level. A generic “Azure costs $X per million” line should not be used for procurement.

Use Azure’s official pricing calculator with the selected region and service tier. If a custom voice or container is involved, evaluate its eligibility and commercial terms as a separate project.

Amazon Polly — Four Engine Classes

Amazon’s pricing page currently distinguishes Standard, Neural, Long-Form, and Generative speech. At the August 1, 2026 review, its listed examples were $4, $16, $100, and $30 per million characters, respectively. Those values are a dated snapshot, not a promise for future billing.

AWS free access can depend on account age, credits, or the selected engine. Use the correct class and include repeated previews. Long-Form and Generative should not be priced using the Standard rate.

PlayHT, Murf, and WellSaid — Production Subscriptions

These products commonly package voice generation with workspace, collaboration, export, or production features. Their public offers can change more quickly than a cloud utility tariff. Check the official checkout screen and capture the plan name, billing cycle, included generation, overage policy, seats, voice cloning access, API availability, and output rights.

For these tools, the value may be editing speed or team workflow rather than the lowest theoretical character price. Run a real project through the trial and count usable final minutes, not only generated minutes.

Speechify — Reading Subscription vs Other Products

Speechify’s main reading product combines document listening, OCR, apps, and other productivity features. That is not directly comparable to a metered developer API. Price the exact product used, confirm whether audio can be exported and published, and distinguish cloud voices from its supported on-device iOS voices.

NaturalReader — Personal and Commercial Are Separate

NaturalReader’s current Personal plans have different Free, Lite, Plus, and Pro voice allowances. Its documentation states that Personal audio is for private use. Public narration belongs in its separately priced Commercial AI Voice Generator. Comparing the personal plan to a commercial API without this boundary understates the required cost.

Self-Hosted Kokoro — No Provider Bill, Not Zero Cost

Kokoro’s model card lists an Apache-2.0 license, but a self-hosted service still requires hardware or rented compute, engineering, observability, security patching, storage, backups, and support. Browser use and a multi-user production server are also different architectures.

Self-hosting can be economical when utilization is stable and the team already operates inference systems. It can be more expensive than an API for sporadic workloads or when engineering time is scarce.

Calculate Cost per Million Characters Correctly

For a character-metered provider, use:

effective cost per 1M = total monthly TTS charges ÷ billable characters × 1,000,000

Do not use plan price alone when the same pool pays for other products. Also document whether spaces, punctuation, SSML tags, and non-Latin characters are counted differently.

For credits or tokens:

1. select one model and output setting;
2. generate a representative 10,000-character batch;
3. record credits or tokens before and after;
4. include discarded takes and retries;
5. calculate the observed unit cost;
6. repeat for a long-form batch.

This produces a project-specific number that can be audited later.

Cost Scenarios Compared

Individual Narrator

A person generating several short scripts manually may prefer OfflineTTS because there is no provider bill and the local UI is sufficient. If the selected device is slow or a premium voice avoids hours of editing, a paid plan can still have the lower total cost.

Application With Variable Traffic

A managed API is often easier when requests come from software and traffic changes month to month. Compare Google, Azure, and AWS using the exact voice class, then include storage, egress, retries, monitoring, and support. OfflineTTS is not a server API substitute.

Long-Form Production Team

For audiobooks, courses, or localization, track usable finished minutes, pronunciation revisions, collaborative editing, and rights—not only input characters. Studio tools or a more controllable voice may reduce labor enough to outweigh higher metered cost.

High-Volume Self-Hosted Service

Model throughput must be measured on the chosen hardware with concurrent jobs. Include idle time, redundancy, maintenance, and engineering. Do not assume a mini PC “pays for itself in a month” without a workload trace and a cloud baseline using the same output requirements.

Verification Rules for the 2026 Price Table

This table was editorially reviewed on August 1, 2026. Official pricing pages were used for ElevenLabs, Google Cloud, Azure, and Amazon Polly; the Kokoro model card was used for the model license. Entries without a stable utility tariff intentionally describe the billing structure instead of repeating an unverified checkout price.

To reproduce an estimate, save a dated copy or screenshot of the official price, record region and currency, identify the exact plan and model, and calculate from a representative batch. Recheck immediately before procurement. Promotions, taxes, app-store pricing, enterprise contracts, free credits, and legacy plans should be listed separately rather than blended into the baseline.

Final Recommendations

• Use OfflineTTS for manual local auditions and exports when its engine selection and device limits fit.
• Use a managed cloud API when application integration, centralized operations, and elastic traffic are primary requirements.
• Pay for a creator platform when its voice, editing, cloning, document, or collaboration feature demonstrably reduces production work.
• Consider self-hosting only after measuring workload, throughput, operating labor, and reliability needs.

The cheapest plan is the one that produces an approved, properly licensed file with the least total labor and operational risk. Start with a representative project, not a provider’s headline unit.