Google Cloud Text-to-Speech Pricing 2026: Voice Models, Free Limits and Cost per Hour

Published on July 15, 2026 by Steven Jones

Google Cloud Text-to-Speech pricing starts at $4 per million characters for Standard and WaveNet voices, rising to $160 per million characters for Studio voices. Chirp 3 HD costs $30 per million characters, whereas Gemini-TTS uses separate text and audio token charges rather than character-based billing.

Those rates are easy to read but awkward to budget. A script is measured in words, a finished voice-over is measured in minutes, and the final invoice is affected by spaces, SSML markup, pronunciation fixes and repeated renders. This guide translates Google’s current rates into cost per 1,000 words, cost per narration hour and cost per accepted output. It also explains when Google is cheaper or more expensive than ElevenLabs, OpenAI, and Azure.

Quick answer: using a working estimate of 6 billable characters per English word and a speaking rate of 150 words per minute, one finished narration hour contains about 54,000 characters. That makes the approximate pre-retry cost $0.22 with Standard or WaveNet, $0.86 with Neural2, $1.62 with Chirp 3 HD and $8.64 with Studio.

Google Cloud Text-to-Speech pricing table for 2026

The table below reflects Google’s published US dollar list prices checked on 15 July 2026. Billing must be enabled even when usage remains inside a free allowance. Accounts billed in another currency can see different SKU prices, taxes and exchange-rate effects.

Voice model or family	Free monthly usage	Price after free usage	Approx. cost per 1,000 words	Approx. cost per narration hour
Standard	4 million characters	$4 per million characters	$0.024	$0.22
WaveNet	4 million characters	$4 per million characters	$0.024	$0.22
Neural2	1 million characters	$16 per million characters	$0.096	$0.86
Polyglot Preview	1 million characters	$16 per million characters	$0.096	$0.86
Chirp 3 HD	1 million characters	$30 per million characters	$0.18	$1.62
Instant Custom Voice	None	$60 per million characters	$0.36	$3.24
Studio	1 million characters	$160 per million characters	$0.96	$8.64
Gemini 2.5 Flash TTS and Flash-Lite Preview TTS	None	$0.50 per million text tokens plus $10 per million audio tokens	About $0.10	About $0.91
Gemini 3.1 Flash TTS Preview and Gemini 2.5 Pro TTS	None	$1 per million text tokens plus $20 per million audio tokens	About $0.20	About $1.81

Calculations assume 6 characters per word and 150 spoken words per minute. Gemini estimates use Google’s rate of 25 audio tokens per second and exclude unusually long style prompts.

Google currently labels WaveNet, Studio, Standard, Neural2 and Polyglot as legacy TTS models. Legacy does not mean unavailable, but it should affect a new implementation decision. A low price is less useful if a team later has to remap hundreds of voice IDs or rebuild a production pipeline around a replacement family.

There is another billing detail worth checking in the Cloud console: Standard and WaveNet appear under the same pricing SKU, while Neural2 and Polyglot also share a SKU. Do not automatically treat matching rows as separate free pools. Budget conservatively as though usage within each shared SKU is combined, then verify the allowance shown in your own billing account.

---

---

What Google counts as a billable character

Google counts the full input string sent for synthesis. Letters and numbers count, but so do spaces, punctuation marks and newline characters. For SSML requests, almost every tag and attribute also counts. The exception listed by Google is the <mark> tag.

• Spaces count: a 10-word sentence contains more than the letters inside its words.
• Punctuation counts: commas, apostrophes, quotation marks and full stops contribute to usage.
• Line breaks count: formatting a script across many short lines slightly increases the total.
• SSML generally counts: tags such as <break time="500ms"/>, <say-as> and <phoneme> add billable characters even though the markup is not spoken aloud.
• Repeated requests count again: regenerating the same paragraph sends the characters through the service another time.

For most English narration, SSML overhead is not the main cost. It becomes noticeable in heavily marked-up scripts containing pronunciation instructions, timing controls and voice changes on nearly every line. At Studio pricing, an extra 10,000 characters of markup costs $1.60. With Standard or WaveNet, the same overhead costs $0.04.

Google also notes that a character is not always the same as a byte. Some writing systems use multiple UTF-8 bytes to represent a single visible character. For Standard and WaveNet billing, Google states that these are charged as one character rather than several bytes. Request size limits can still be byte-based, so a script may hit an API limit before it reaches the character count expected for English text.

The authoritative rates and billing definition are available in the Google Cloud Text-to-Speech pricing documentation.

How to calculate Google TTS cost per 1,000 words

A practical English-language estimate is 6 billable characters per word once letters, spaces and ordinary punctuation are included. It is not exact. Technical scripts with URLs, code, serial numbers, or SSML can be higher, while scripts with short conversational words can be lower.

Use this formula for character-priced voices:

Estimated cost = word count x average characters per word x price per character

For a 1,000-word script at 6,000 characters:

• Standard or WaveNet: 6,000 x $0.000004 = $0.024
• Neural2 or Polyglot: 6,000 x $0.000016 = $0.096
• Chirp 3 HD: 6,000 x $0.00003 = $0.18
• Instant Custom Voice: 6,000 x $0.00006 = $0.36
• Studio: 6,000 x $0.00016 = $0.96

The free allowance should be applied at the monthly account or SKU level, not deducted from every individual job. A 1,000-word Chirp render may show no incremental charge while the account is inside its first million characters, but it has still consumed roughly 0.6% of that allowance.

How many characters are in one hour of narration?

At 150 words per minute, one hour contains about 9,000 spoken words. Multiplying that by 6 characters per word gives a working estimate of 54,000 billable characters per hour.

Monthly free allowance	Approx. words	Approx. finished narration
4 million characters	666,667 words	74.1 hours
1 million characters	166,667 words	18.5 hours
500,000 characters	83,333 words	9.3 hours

The 500,000-character row is included for easier comparison with Azure’s free neural TTS allowance.

Speaking pace changes duration without necessarily changing character usage. A slower read of the same character-priced script does not cost more simply because the audio lasts longer. Gemini-TTS is different because audio output tokens are tied to duration. Slowing the delivery can therefore increase the output token cost even when the spoken text is unchanged.

Gemini-TTS pricing needs a different calculator

Gemini-TTS cannot be compared with Chirp or Neural2 by multiplying characters. Google charges for text input tokens and generated audio tokens, with 25 audio tokens corresponding to 1 second of output. The audio charge dominates ordinary narration costs.

At current rates, Gemini 2.5 Flash TTS and Flash-Lite Preview TTS cost about $0.015 per minute generated before the small text-input charge. Gemini 3.1 Flash TTS Preview and Gemini 2.5 Pro TTS cost about $0.03 per generated minute. A one-hour output is therefore roughly $0.91 or $1.81, respectively, under the assumptions used here.

The pricing model has an operational consequence. With character billing, tighter pacing can increase the amount of spoken content produced for the same charge. With audio-token billing, long pauses, slower delivery and extended performance directions can increase cost. The expressive version may sound better, but it should be budgeted like generated audio rather than like converted text.

Regeneration rate is the hidden number in a narration budget

Published API rates describe generated output, not accepted output. Names, acronyms, product codes, dates and sentence emphasis often need correction. If a team regenerates an entire five-minute section to fix one word, the cost of the accepted narration includes both versions.

A better production metric is:

Cost per accepted hour = base generation cost / acceptance rate

Voice family	90% accepted first time	75% accepted first time	50% accepted first time
Standard or WaveNet	$0.24	$0.29	$0.43
Neural2 or Polyglot	$0.96	$1.15	$1.73
Chirp 3 HD	$1.80	$2.16	$3.24
Studio	$9.60	$11.52	$17.28

This models repeated generation only. Human editing, storage, application hosting and engineering time are separate.

Real-world voice workflows repeatedly expose the same failure mode: long narration is divided into chunks to satisfy request limits, then adjacent clips differ slightly in pace, energy or pronunciation. Regenerating whole chunks to recover continuity can cost more than fixing isolated words. This is especially relevant to prompt-steered models, where identical voice selection does not guarantee an identical performance across independent requests.

Why the cheapest voice can cost more after editing

Standard and WaveNet are extremely cheap on the API bill. They are not automatically the lowest-cost production choice. A voice that needs more manual cuts, pronunciation patches, or cadence correction transfers costs from Google Cloud to the editor.

Suppose one hour of Chirp 3 HD costs $1.62 and saves 20 minutes of editing compared with a $0.22 Standard render. At an internal editing cost of $30 per hour, the saved labour is worth $10. The $1.40 API premium is trivial compared to the production savings. The opposite can also happen: if a premium voice still needs the same editing, its higher rate buys no economic benefit.

The sensible test is not to ask which demo sounds most human. Render the same difficult five-minute script with each candidate, then record:

• accepted minutes from the first render
• number of pronunciation corrections
• time spent repairing pacing and emphasis
• consistency between independently generated chunks
• total API cost for every attempt
• editing minutes required before publication

This results in a cost per publishable minute rather than per API request. It is the only comparison that remains useful once a workflow moves beyond occasional prototypes.

Google Cloud TTS versus ElevenLabs, OpenAI and Azure

Provider	Typical billing basis	Approx. raw narration cost	Where it makes sense	Main budget risk
Google Cloud TTS	Characters for most voices; tokens for Gemini-TTS	$0.22 to $8.64 per hour across the main prebuilt character-priced families	API-led products, high-volume narration and teams already using Google Cloud	Voice-family complexity, legacy model planning and regeneration across chunks
ElevenLabs	Characters or monthly credits	API list pricing is roughly $0.05 to $0.10 per 1,000 characters	Creator workflows, voice libraries, cloning and projects where a production interface saves editing time	Credits are consumed during iteration, and shared credits can be used by other audio features
OpenAI	Text input and audio output tokens	GPT-4o mini TTS is approximately $0.90 per generated hour at ordinary pacing	Prompt-controlled delivery and applications already built around OpenAI APIs	Slower or longer output increases audio-token usage
Azure AI Speech	Characters	About $0.81 per hour at $15 per million neural TTS characters	Microsoft-centred enterprise deployments, regional controls and established Azure estates	Voice tier, custom voice and hosting choices can create separate charges

Google is the raw-cost leader for Standard and WaveNet, and its current free allowances are generous for modest narration volumes. Azure’s neural rate sits close to Google’s Neural2. OpenAI’s duration-linked economics are competitive for steerable speech. ElevenLabs is materially more expensive by character at API list price, but its creator tools can still be cheaper overall when they reduce integration and editing work.

For a broader view of voice, music and audio production products, see our guide to the best AI audio generation tools. Our Fish Audio review also examines a creator-focused alternative where voice workflow and model choice matter more than Cloud infrastructure.

The break-even point between Google API narration and a creator subscription

Comparing a metered API with a creator subscription requires two break-even calculations. The first compares the fixed monthly fee with Google’s API bill. The second values the production features included in the subscription.

Using $22 per month as a reference subscription price, Google’s bill reaches the same level at approximately:

Google voice family	Approx. generated hours before the monthly bill reaches $22
Standard or WaveNet	176 hours, including the estimated 74-hour free allowance
Neural2 or Polyglot	44 hours, including the estimated 18.5-hour free allowance
Chirp 3 HD	32 hours, including the estimated 18.5-hour free allowance
Studio	21 hours, including the estimated 18.5-hour free allowance
Instant Custom Voice	6.8 hours because there is no free allowance
Gemini 2.5 Flash TTS	About 24 hours because there is no free allowance

These are generated-hour comparisons, not feature or capacity equivalents. Free allowances reset monthly and may be shared within a billing SKU.

The table makes Google look unbeatable, but a creator subscription is not merely a bag of characters. It may include a browser editor, project management, voice cloning, pronunciation controls, reusable voice settings and easier export. If a $22 subscription saves 44 minutes of work per month for someone whose time costs $30 per hour, it has already recovered its fee.

The practical break-even point is therefore much lower for a solo creator who does not want to build an interface, and much higher for a developer who already has storage, rendering, review and retry controls in place. API pricing wins on unit cost. Subscriptions can win on avoided labour.

A production workflow that prevents avoidable TTS charges

Most waste is introduced before the synthesis request is made. The following workflow keeps retries local and makes the eventual cloud bill explainable.

1. Normalise the script first. Expand abbreviations, decide how dates and currencies should be spoken, and remove text that should not reach the voice model.
2. Build a pronunciation test block. Render names, acronyms, URLs, model numbers and brand terms before generating the full script.
3. Count the actual payload. Measure characters after template insertion and SSML generation, not the clean source document.
4. Chunk by edit boundary. Use paragraphs or scenes that can be replaced independently. Tiny chunks create assembly work; huge chunks make one mistake expensive.
5. Cache accepted audio. Store a hash of the text, voice, speaking settings and model. Do not regenerate a clip that has not changed.
6. Log the retry reason. Separate pronunciation, style, service error and script-change retries. Only the first two indicate a voice-quality problem.
7. Track accepted-output cost. Divide total monthly synthesis spend by published minutes, not by requested characters alone.
8. Set a billing alert. Free usage is not a hard cap. Google automatically charges beyond the allowance once billing is enabled.

For repeated production, keep voice settings and text normalisation under version control. A model update, a renamed voice, or an altered default can otherwise change the sound of future episodes without an obvious code change in the application.

Costs outside the Text-to-Speech line item

The synthesis rate is only one part of a deployed system. Long-form generation can require Cloud Storage, and an application may also use serverless compute, queues, databases, logging and network transfer. Google bills those services separately.

• Cloud Storage: source scripts, intermediate WAV files and final assets can accumulate quickly when every retry is retained.
• Compute: functions or containers may handle text normalisation, API requests, audio stitching and format conversion.
• Network transfer: serving large audio files outside Google Cloud can add egress charges.
• Monitoring: detailed logs help explain cost but can themselves become billable at scale.
• Engineering: authentication, retry handling, quota management, review screens and pronunciation tooling all require implementation time.
• Post-production: loudness normalisation, silence trimming, music mixing and quality review are not included in the TTS rate.

A small narration project may never notice these extras. A product serving personalised audio to thousands of users can spend more on surrounding infrastructure and delivery than on Standard voice synthesis.

Which Google Cloud voice tier offers the best value?

Standard or WaveNet is the cost-first choice for utility speech, accessibility playback and high-volume content where editorial performance is not the main product. The current $4-per-million-character rate makes both difficult to beat on a raw cost basis.

Neural2 is the safer default for a conventional, new-character-priced integration. Its $16 rate remains low, its free allowance covers roughly 18.5 narration hours under our assumptions, and its general-purpose positioning avoids paying the Studio premium before the workflow proves it needs it.

Chirp 3 HD fits conversational products and projects that need Google’s newer voice family, streaming support or more modern controls. At about $1.62 per generated hour, the raw price is still small compared with editing and application development.

Gemini-TTS is the more logical evaluation route when delivery style, multi-speaker output and natural-language direction are central. Budget it by the generated duration, test consistency across chunks, and avoid assuming that prompt-steered performance will reproduce identically on every request.

The studio needs a clear justification for quality or workflow. Its $160-per-million-character rate is 10 times Neural2 and more than five times Chirp 3 HD. It can still be economical compared to human narration, but it should win a controlled acceptance-rate test rather than be selected solely because it is the most expensive prebuilt tier.

Google Cloud Text-to-Speech pricing FAQs

Is Google Cloud Text-to-Speech free?

Several voice families have monthly free usage, but billing must still be enabled. Standard and WaveNet list 4 million free characters, while Neural2, Polyglot, Chirp 3 HD and Studio list 1 million. Instant Custom Voice and Gemini-TTS do not currently list a free usage allowance.

Do spaces and punctuation count towards Google TTS pricing?

Yes. Google counts spaces, punctuation and newline characters in the input string. This is why a word-count-only estimate should include an average-character multiplier rather than counting letters alone.

Does SSML increase the bill?

Usually. Google counts SSML tags and attributes as characters, apart from the <mark> tag. Light markup has little effect on low-cost voices, but dense pronunciation and timing markup can become noticeable with Studio or at high volume.

How much does one hour of Google TTS narration cost?

Using 54,000 characters per finished hour, the approximate pre-retry cost is $0.22 for Standard or WaveNet, $0.86 for Neural2, $1.62 for Chirp 3 HD, $3.24 for Instant Custom Voice and $8.64 for Studio. Gemini 2.5 Flash TTS is approximately $0.91 per generated hour under ordinary pacing.

Is Google Cloud TTS cheaper than ElevenLabs?

Google is substantially cheaper on raw character generation for most prebuilt voices. ElevenLabs can still offer better production economics for creators who value its interface, voice library, cloning and editing workflow. Compare cost per publishable minute, not API price alone.

Why was the actual Google TTS bill higher than the script estimate?

The common causes are SSML characters, repeated renders, regenerated chunks, a different voice SKU, usage from another application in the same project, and surrounding Google Cloud services. Count the final request payload and reconcile it against model-specific usage rather than relying on the source document’s word count.

The practical pricing verdict

Google Cloud Text-to-Speech is inexpensive enough that voice selection should not be driven solely by the lowest rate. Standard and WaveNet can produce an hour for about $0.22, but even Chirp 3 HD is only around $1.62 before retries. One avoided editing pass can outweigh that difference many times over.

Start with Neural2 for predictable general-purpose character billing, Chirp 3 HD for newer conversational workflows, and Gemini-TTS when prompt-controlled delivery is central to the product. Run a difficult five-minute acceptance test before committing. The winning model is the one that produces the lowest cost per approved minute while remaining consistent enough for the next hundred jobs, not the one with the cheapest line in the pricing table.