Last updated: August 2026
Which text-to-speech API should you actually integrate?
It depends on your stack, your volume, and whether latency is the deciding factor or just one input among several.
This is the general-purpose answer to “which text-to-speech API should I use,” not the latency-obsessed one. If you’re building a live voice agent, a phone bot, anything where a person is waiting on the other end of the line in real time, our guide to AI voice generators for real-time apps and voice agents already covers that specific comparison (Cartesia, Inworld, Hume, and ElevenLabs Flash) in depth, and repeating it here would just be the same content under a different title. This guide is for the more common case: you need to turn text into audio somewhere in your product, an in-app narrator, notification read-aloud, accessibility feature, IVR system, batch-generated audiobook chapter, and you want to know what it actually costs and which provider fits your existing stack.
What actually differs between text-to-speech APIs?
Every provider in this space will tell you their voices sound the most natural, so that’s not a useful first filter. Four things actually separate these APIs in practice:
- Pricing tiers, and they’re rarely simple. Almost every provider charges different rates for different voice quality levels (a “Standard” or “WaveNet” tier vs. a premium “Studio” or “Generative” tier), and the price gap between the cheapest and most expensive tier from the same vendor can be 20-40x.
- Whether you’re already paying for the parent cloud platform. If your infrastructure is already on AWS, GCP, or Azure, using that vendor’s TTS API means one bill, one IAM setup, and no new vendor relationship to manage. That convenience is worth real money even if a competitor’s voices sound marginally better.
- Free tier size and what happens after it. Free tiers here range from 500,000 characters a month (Azure) to 4 million (Google Standard tier), and what counts toward the free tier (does the premium voice tier burn through it faster?) varies by vendor.
- Whether you need bundled speech-to-text too. Some vendors (Deepgram in particular) are built around pairing TTS with STT and orchestration in one bundle, which matters if you’re building a full voice pipeline rather than just generating audio from a script.
Best text-to-speech APIs for developers, provider by provider
Google Cloud Text-to-Speech: The widest voice and language selection of any provider here, 700+ voices across 40+ languages and 100+ locales, spread across four pricing tiers. Standard and WaveNet voices run $4 per million characters, Neural2 is $16, the newer Chirp 3 HD voice (released March 2026) is $30, and the premium Studio tier is $160. The free tier is generous relative to competitors: 4 million Standard characters a month, plus 1 million each of WaveNet, Neural2, and Chirp 3. Best default if language coverage matters more than any other single factor.
Amazon Polly: AWS’s native TTS service, priced almost identically in structure to Google’s: Standard voices at $4 per million characters, Neural at $16, Generative at $30, and a Long-Form tier (built for narrating longer content with more natural pacing) at $100. The free tier covers 1 million Neural characters a month for the first 12 months. Amazon expanded Generative voice availability to 8 regions as of March 2026, including Europe (London) and Canada (Central). The obvious choice if your infrastructure is already AWS-native and you don’t want a second cloud vendor relationship just for audio.
Microsoft Azure AI Speech: Standard Neural voices are $16 per million characters, with newer Neural HD voices at $22, down from $30 as of a March 2026 price cut. Custom Neural voices (training a voice on your own samples) run $24-48 per million characters. The free tier is smaller than Google’s or Amazon’s at 500,000 characters a month, but Azure’s commitment tiers reward scale aggressively: $960/month gets you 80 million characters (effectively $12/1M), and the largest tier ($15,000/month for 2 billion characters) works out to $7.50 per million, the cheapest effective rate of any provider covered here at genuine enterprise scale. Makes the most sense for teams already standardized on Microsoft’s cloud or needing custom voice training.
OpenAI TTS: The simplest integration path if your product already calls OpenAI’s API for anything else, since it’s the same account, same billing, same SDK pattern. Two flat-rate models: tts-1 at $15 per million characters and tts-1-hd at $30. A newer option, gpt-4o-mini-tts, prices differently, by token rather than by character, at $0.60 per million input text tokens plus $12 per million audio output tokens, and supports more voices (13 vs. 9 on the older models) along with steerable delivery prompts. Not the cheapest option at scale, but the lowest-friction one if OpenAI is already in your stack.
Deepgram (Aura-2): Positioned differently from the four above: Deepgram is built around pairing TTS with its own speech-to-text and orchestration, not just generating audio in isolation. Aura-2 costs $0.030 per thousand characters on pay-as-you-go pricing (roughly $30 per million), dropping to $0.027 per thousand at the Growth tier, which requires a $4,000+ prepaid commitment. Latency is genuinely strong for a general-purpose API, 90ms in optimized conditions, sub-200ms baseline, though it supports fewer languages than Google or Amazon (7, versus 40+). There’s also a bundled Voice Agent API combining STT, TTS, and orchestration at $4.50 an hour, worth knowing about if you’re building a full conversational pipeline rather than shopping for TTS alone, though our dedicated real-time voice generator guide covers that specific comparison against Cartesia and Inworld in more depth.
How do the major TTS APIs compare on price?
| Provider | Entry tier | Premium tier | Free tier | Best for |
|---|---|---|---|---|
| Google Cloud TTS | $4/1M (Standard) | $160/1M (Studio) | 4M chars/mo | Widest language coverage |
| Amazon Polly | $4/1M (Standard) | $100/1M (Long-Form) | 1M chars/mo (12 months) | AWS-native stacks |
| Azure AI Speech | $16/1M (Neural) | $48/1M (Custom) | 500K chars/mo | Cheapest at 2B+ char scale ($7.50/1M) |
| OpenAI TTS | $15/1M (tts-1) | $30/1M (tts-1-hd) | None published | Teams already on OpenAI’s API |
| Deepgram Aura-2 | ~$30/1M (pay-as-you-go) | ~$27/1M (Growth, $4K+ commit) | $200 credit | Bundled STT+TTS pipelines |
These are list prices for the base tiers described above; every vendor here also offers custom enterprise pricing at high volume that isn’t published, so treat this table as a starting comparison, not a final quote. Confirm current rates directly with each vendor before committing, pricing on this list moved twice in the sources we checked (Azure’s Neural HD cut, Google’s Chirp 3 HD launch) within the past several months alone.
What does a real project actually cost on each provider?
List prices per million characters are hard to picture until they’re attached to something concrete. Take a mid-size audiobook, roughly 90,000 words, which works out to about 500,000 characters including spaces, at the entry-quality tier from each provider:
- Google Cloud TTS (Standard/WaveNet, $4/1M): about $2 for the full book, and it fits entirely inside the 4 million character monthly free tier if this were your only usage that month.
- Amazon Polly (Standard, $4/1M): also about $2, though Polly’s more natural-sounding Neural tier at $16/1M would run closer to $8 for the same book.
- Azure AI Speech (Neural, $16/1M): around $8, using up most of the 500,000-character free tier in one project.
- OpenAI TTS (tts-1, $15/1M): roughly $7.50, with no published free tier to offset it.
- Deepgram Aura-2 (~$30/1M): around $15, the most expensive of the five at this content type and volume, though Deepgram isn’t really positioned to compete on narration-style pricing in the first place.
The gap looks small at book-length volume, a few dollars either way, but multiply by hundreds or thousands of pieces of content a month (a news reader app, a customer notification system, an e-learning platform generating per-lesson audio) and the same percentage gaps turn into real budget line items. Model your actual monthly character volume against these tiers before assuming any one provider is “the cheap one” in general.
A higher-volume comparison tells a different story, and it’s worth being careful with the math rather than assuming a commitment tier automatically saves money. Take a notification-reading feature inside a consumer app generating 50 million characters a month. At that volume, Google Standard and Amazon Polly Standard both land around $200/month at list price, and Azure’s Neural tier at list price would run $800/month. Azure’s commitment tiers only pay off once actual usage is close to the tier’s full allowance: the $960/month tier is priced for 80 million characters, so a team using only 50 million against that commitment would be paying roughly $19.20 per million characters effectively, worse than list price, not better. Commitment pricing is a real discount, but only once your real monthly volume genuinely approaches (not just qualifies loosely for) the tier you’re paying for.
How much do SDK and language support actually differ?
All five providers offer REST APIs directly, so nothing here is a hard blocker, but the maturity of official client libraries varies. Google Cloud and Amazon both ship full first-party SDKs across the usual list (Python, Node.js, Java, Go, C#, Ruby, PHP), consistent with their broader cloud platforms, since TTS is just one more service alongside everything else in those SDKs. Azure’s Speech SDK is similarly mature but is a separate, dedicated SDK from Azure’s general-purpose cloud SDK, worth knowing before assuming it will feel identical to other Azure service integrations. OpenAI’s TTS endpoint uses the same lightweight client library as the rest of its API, which is the fastest to get a first request working if you’ve already integrated any other OpenAI endpoint. Deepgram provides dedicated SDKs as well, with particular attention to streaming use cases given its speech-to-text roots, so its client libraries tend to handle WebSocket-based streaming audio more natively than a pure batch-oriented TTS call.
What changed with these APIs in 2026?
Three concrete moves happened within the last several months across these providers. Google shipped Chirp 3 HD in March 2026 at $30 per million characters, sitting between its existing Neural2 and Studio tiers, an attempt to offer near-premium quality without premium pricing. Azure cut its Neural HD voice pricing from $30 to $22 per million characters in the same month, a direct response to competitive pressure on quality-tier pricing across the category. And Amazon expanded Generative voice availability to two new regions (Europe/London and Canada/Central), bringing total Generative coverage to 8 regions, which matters if data residency requirements have been the blocker keeping a team on an older voice tier.
What should you actually test before committing to one?
List pricing tells you what a provider costs in theory. Before locking in an integration, run three checks against your own use case rather than trusting the comparison table above at face value:
- Generate your actual scripts, not demo text. Voice quality perception varies a lot by content type, a provider that sounds great reading a marketing script can sound flat reading technical documentation or numbers-heavy content.
- Check what counts against your free tier. Some vendors count every character including whitespace and SSML markup tags toward billing, which can meaningfully inflate real usage above what a naive character count of your script suggests.
- Confirm the voice tier your quote actually assumed. A team comparing “Google vs. Amazon” without specifying voice tier is often comparing Google’s $16 Neural2 tier against Amazon’s $4 Standard tier, a 4x price difference that has nothing to do with which vendor is actually cheaper at matched quality.
Sources
- Google Cloud, official Text-to-Speech pricing documentation, 2026, retrieved 2026-08-03
- Amazon Web Services, official Amazon Polly pricing documentation, 2026, retrieved 2026-08-03
- Microsoft Azure, official AI Speech pricing documentation, 2026, retrieved 2026-08-03
- OpenAI, official API pricing documentation (TTS models), 2026, retrieved 2026-08-03
- Deepgram, official Aura-2 and Voice Agent API pricing documentation, 2026, retrieved 2026-08-03
Pricing for cloud TTS APIs changes frequently and varies by region, commitment tier, and negotiated enterprise rates. We verified the figures above against each vendor’s own published pricing pages at the time of writing; confirm current rates directly before budgeting a production integration.
Frequently Asked Questions
What’s the cheapest text-to-speech API for developers?
At entry-level voice quality, Google Cloud TTS and Amazon Polly are tied at $4 per million characters for their Standard tiers. At genuine enterprise scale (2 billion+ characters), Azure’s commitment pricing drops to $7.50 per million, the lowest effective rate among the providers covered here.
Should I use my cloud provider’s TTS API or a specialized vendor?
If your infrastructure is already on AWS, Google Cloud, or Azure, using that vendor’s native TTS API avoids a second billing relationship and separate IAM setup, which is a real operational simplification even if a specialized vendor’s voices sound marginally better.
Is OpenAI’s TTS API good for production use?
Yes, particularly if your product already integrates OpenAI’s API for other functionality. It isn’t the cheapest option at scale ($15-30 per million characters), but it has the lowest integration friction for teams already inside that ecosystem.
Which TTS API should I use for a live voice agent instead of general narration?
None of the five covered in this guide are optimized specifically for that. See our guide to AI voice generators for real-time apps and voice agents, which covers Cartesia, Inworld, Hume, and ElevenLabs Flash specifically for latency-sensitive conversational use.
Does Deepgram only do speech-to-text?
No. Deepgram is best known for speech-to-text, but its Aura-2 product is a full text-to-speech API, and it also offers a bundled Voice Agent API combining STT, TTS, and orchestration for teams building complete conversational pipelines.
Why do prices vary so much between voice tiers from the same provider?
Higher tiers (Studio, Generative, Neural HD) typically use larger, more expressive models that cost more to run, and are priced accordingly, sometimes 10-40x the entry tier from the same vendor. Confirm which tier a price quote assumes before comparing providers directly.
Do these APIs support SSML for controlling pronunciation and pacing?
Google Cloud TTS, Amazon Polly, and Azure AI Speech all support SSML (Speech Synthesis Markup Language) for controlling pauses, emphasis, and pronunciation of specific words. OpenAI’s TTS API does not use SSML, instead relying on plain-text prompting for delivery control on its newer gpt-4o-mini-tts model. If precise pronunciation control (acronyms, names, technical terms) is a requirement, confirm SSML support specifically rather than assuming it’s universal across providers.
Pricing, free tier limits, and voice availability for the APIs above change frequently. Always confirm current figures directly with each vendor before committing to a production integration.
Richard Johnson
Richard Johnson is an AI specialist with over five years of experience guiding large organizations through AI adoption, across more than 100 customers. He founded CognitiveFuture to research and compare AI tools across design, development, writing, research, voice and business, cutting a crowded, fast-moving market down to the right choice for the job in front of you.
