Best Text-to-Speech API for Indie Developers (2026, Priced)

Short answer: Start on Google Cloud Text-to-Speech or Amazon Polly if you want the lowest bill, because both have free monthly character allowances and paid tiers from $4 per 1M characters. Use ElevenLabs when the voice itself is the product (narration, characters, cloning), Deepgram Aura-2 or Cartesia for real-time voice agents, and self-hosted Kokoro if you want no per-character fee at all. Avoid starting anything new on OpenAI's speech endpoint models: they shut down on 6 January 2027.

Every price, limit and licence below comes from the vendor's own page or model card, checked on 12 Nov 2026. Advice is ours and marked "Our read".

TL;DR#

• Cheapest managed voices: Google Standard and WaveNet voices at $4 per 1M characters with the first 4M free (Google Cloud TTS pricing); Amazon Polly Standard at $4 per 1M with 5M free per month (Polly pricing).
• Best value natural voices: Google Chirp 3: HD at $30 per 1M after the first 1M characters; Polly Neural at $16 per 1M.
• Best quality and cloning: ElevenLabs at $0.04 (Flash/Turbo) to $0.08 (Multilingual v2, v3) per 1,000 characters (ElevenLabs API pricing).
• Voice agents: Deepgram Aura-2 at $0.030 per 1,000 characters (Deepgram pricing) or Cartesia Sonic from $5/month (Cartesia pricing).
• Free and self-hosted: Kokoro-82M, Apache 2.0 (model card).
• Avoid for new builds: OpenAI tts-1, tts-1-hd and gpt-4o-mini-tts, removed 6 January 2027 (OpenAI deprecations).

Disclosure: ZeroRaised is a launch site for bootstrapped products. We have no commercial relationship with any vendor in this post and earn nothing if you sign up. Facts come from the linked sources as checked on 12 Nov 2026.

What indie developers actually need from a TTS API#

Most indie use cases are either batch audio (read-aloud, audiobook chapters, voiceover for a product demo video), where cost and quality matter and latency does not, or real-time voice, where the reply must start speaking fast. For the second, our guide on how to build an AI voice agent covers the full pipeline; this post only covers speech out.

Three things matter more than the headline price: whether the free tier allows commercial use, whether billing is per character or per token, and whether the model will still exist next year.

Text-to-speech API prices compared#

Sources: Google, Polly, ElevenLabs, Deepgram, Cartesia, OpenAI, Azure, Kokoro.

Google says you are charged "if your usage exceeds the number of free characters allowed per month", and billing must be enabled to use the API at all. Azure's paid per-character prices did not render on its page when we checked, so only its free tier is listed.

The options, one by one#

Google Cloud Text-to-Speech: best free tier for real products#

Google's pricing page lists Chirp 3: HD as a current model: first 1M characters free, then "US$30 per 1 million characters". Standard and WaveNet sit under legacy models at $4 per 1M with 0 to 4M characters free. Gemini TTS models are billed per token instead (Gemini 2.5 Pro TTS lists $1.00 per 1M text input tokens and $20.00 per 1M audio output tokens) with no free usage.

Our read: for a side project that reads articles aloud or voices notifications, Chirp 3: HD's free allowance will often cover you entirely. The cost is setup: a billing account and a service account before you hear a word.

Amazon Polly: cheapest at scale on Standard voices#

Polly is the boring, reliable option. Standard voices are $4 per 1M characters with 5M free per month, and the page shows no time limit on that Standard allowance. Neural voices are $16 per 1M, with 1M free each month for the first 12 months. Generative voices are $30 per 1M and Long-Form voices $100 per 1M.

Our read: Polly Standard voices sound dated next to ElevenLabs or Chirp 3: HD, but for accessibility features, IVR prompts or high-volume utility speech, $4 per 1M is hard to beat. If you already run on AWS, it is the least friction.

ElevenLabs: best quality, cloning and expressive voices#

ElevenLabs' API pricing page lists per-character rates that are "the same on every plan": $0.04 per 1,000 characters for Flash/Turbo and $0.08 for Multilingual v2 and v3. Plans bundle characters:

On the main pricing page, Starter is the first plan that lists "Commercial License" and "Instant Voice Cloning". Creator adds Professional Voice Cloning.

Our read: if listeners notice the voice (narration, game characters), ElevenLabs is the default. Its Multilingual models are among the priciest here, so use Flash for conversational or high-volume work.

Deepgram Aura-2: priced for voice agents#

Deepgram lists Aura-2 at $0.030 per 1,000 characters pay as you go ($0.027 on Growth) and the older Aura-1 at $0.0150. New accounts get a $200 credit. That credit is the main attraction for an indie build: if you also use Deepgram for speech-to-text, one account covers both ends of a voice agent.

Cartesia Sonic: low-latency voices on a subscription#

Cartesia bills Sonic at "1 credit per character", or roughly 750 to 800 credits per minute of audio. Free gives 20K credits a month. Pro is $5/month for 100K credits and includes a commercial use licence and instant voice cloning. Startup is $49/month for 1.25M credits and Scale $299/month for 8M credits.

Our read: Pro at $5 is the cheapest paid plan in this list that includes both a commercial licence and cloning, which makes it a sensible test bed for a voice agent prototype.

OpenAI TTS: deprecated, plan around it#

OpenAI still lists tts-1 at $15 per 1M characters, tts-1-hd at $30 per 1M, and gpt-4o-mini-tts at $0.60 per 1M text input tokens and $12.00 per 1M audio output tokens (OpenAI pricing). But its deprecations page says all three "will be removed from the API on January 6, 2027" and tells developers to "Migrate to gpt-realtime-2.1-mini before the shutdown date."

That replacement is not a drop-in swap. Its model page lists the speech generation endpoint (v1/audio/speech) as "Not supported"; audio comes through the Realtime API at $20 per 1M audio output tokens. OpenAI's TTS guide also notes that its usage policies "require you to provide a clear disclosure to end users that the TTS voice they are hearing is AI-generated."

Our read: if you already pick models with our LLM API guide, the lesson is the same here: keep TTS behind a thin interface in your code so a deprecation costs you an afternoon, not a rewrite.

Self-hosting: Kokoro, Piper and the XTTS licence trap#

• Kokoro-82M is "an open-weight TTS model with 82 million parameters" under Apache 2.0. Its model card says it "can be deployed anywhere from production environments to personal projects." This is the safest open-weight pick for a commercial product.
• Piper is a fast local TTS engine. The original repo was archived on 6 Oct 2025 and development moved to OHF-Voice/piper1-gpl, which is GPL-3.0. Voices have a separate licensing document, so check the licence of each voice you ship.
• Coqui XTTS-v2 can clone voices from short samples, but its Coqui Public Model License grants rights "for any non-commercial purpose", and only "so far as you do not receive any direct or indirect payment arising from the use of the model or its output." Do not put it in a paid product.

Our read: self-hosting is cheap only if you already have a server with spare capacity. A GPU box that sits idle most of the month will cost more than Google's free tier.

Worked example: 1M characters a month#

Illustrative arithmetic on list prices, ignoring tax, with each vendor's free allowance applied where it has one. Using Cartesia's rough rate of 750 to 800 credits (characters) per minute, 1M characters is about 1,250 to 1,330 minutes of audio.

Our read: the spread is real. At 1M characters a month the cheapest managed options cost nothing and the premium ones cost $40 to $99. That gap matters at indie scale only if your margin per user is thin. If you resell voice minutes, model the cost per user the way our LLM cost guide suggests: cache repeated phrases, generate once and store the audio file, and only use the premium voice where users hear the difference.

Voice cloning: read the vendor's consent rules#

Cloning is where indie apps get into trouble. Use only the vendor's own rules: OpenAI requires custom voices to be created "from a speaker's consent recording and matching audio sample" through an approval process (TTS guide). ElevenLabs and Cartesia gate cloning to paid plans. Our read: whichever vendor you use, only clone a voice you own or have written permission to use, and keep that permission on file.

Our picks by situation#

Frequently asked questions#

What is the cheapest text-to-speech API for indie developers?#

On list prices checked 12 Nov 2026, Google Cloud TTS Standard and WaveNet voices and Amazon Polly Standard voices cost $4 per 1M characters, with free allowances of 4M (Google) and 5M (Polly) characters. For better-sounding voices, Google Chirp 3: HD includes 1M free characters, then $30 per 1M. Self-hosting Kokoro-82M has no licence fee, but you pay for your own compute.

Is ElevenLabs free for commercial use?#

Not on the free plan as listed. ElevenLabs' pricing page shows Starter ($6/month) as the first plan that includes a Commercial License and Instant Voice Cloning. Creator ($22/month) is the first with Professional Voice Cloning. Check the current terms before you ship.

Should I still use OpenAI's text-to-speech API?#

Not for a new project. OpenAI's deprecations page says tts-1, tts-1-hd and gpt-4o-mini-tts will be removed from the API on 6 January 2027, and recommends migrating to gpt-realtime-2.1-mini, which runs on the Realtime API rather than the speech endpoint. OpenAI's usage policies also require you to tell end users that a TTS voice is AI-generated.

Which TTS API is best for a real-time voice agent?#

Our read: Deepgram Aura-2 ($0.030 per 1,000 characters pay as you go, with $200 in free credit) and Cartesia Sonic (from $5/month for 100K credits, 1 credit per character) are both built for conversational, low-latency use. ElevenLabs Flash at $0.04 per 1,000 characters is the other common choice. Test latency from your own region before committing.

Can I self-host a text-to-speech model commercially?#

Yes, with the right licence. Kokoro-82M is Apache 2.0 and its model card says it can be deployed in production. Piper's current repo (piper1-gpl) is GPL-3.0, and individual voices have their own licences. Coqui XTTS-v2 uses the Coqui Public Model License, which grants rights only for non-commercial purposes.

Ship the voice, then launch it#

Once your app can talk, run through our product launch checklist for bootstrapped makers and submit it to ZeroRaised, the launch site for products built without VC. For more priced guides like this one, join the ZeroRaised newsletter.

Sources#

All last checked 12 Nov 2026.

• Google Cloud Text-to-Speech pricing
• Amazon Polly pricing
• ElevenLabs API pricing
• ElevenLabs pricing
• Deepgram pricing
• Cartesia pricing
• OpenAI API pricing
• OpenAI deprecations
• OpenAI gpt-realtime-2.1-mini model page
• OpenAI text-to-speech guide
• Azure AI Speech pricing
• Kokoro-82M model card
• Piper (archived repo) and piper1-gpl
• Coqui XTTS-v2 licence