The 8 best text-to-speech APIs for AI voice agents in 2026

A text-to-speech (TTS) API converts written text into spoken audio over a REST or WebSocket endpoint. For an AI voice agent, the right one comes down to four things: latency, telephony output, language coverage, and cost at real volume.

Every TTS vendor sounds flawless in a 20-word demo. The one worth building on is the one still holding up on turn 400 of a live conversation, when response time has to stay under the point a listener notices, the audio has to arrive in the format your delivery path needs (8kHz mulaw for a phone call, low-latency streaming for a browser or in-app agent), and the bill has to survive the jump from 500K to 5M characters a month.

The best TTS APIs in 2026:

1. Telnyx: best for production phone agents and regulated environments, with one API across many voices
2. Cartesia: best for the lowest time-to-first-audio
3. Inworld: best independently benchmarked quality per dollar
4. Rime: best low-cost telephony voice with self-host
5. Deepgram Aura: best for consistent agent voices with native telephony output
6. Fish Audio: best broad-language cloning with an open-weight fallback
7. ElevenLabs: best raw voice realism and cloning
8. Google Cloud: best for GCP-native apps and batch narration

TTS API architecture, or the structural design of the system, is another factor that influences latency and streaming capability. The engineering decisions underneath the API matter tremendously.

A live call must receive audio in the correct format while the application is still generating text. A narration pipeline can wait for a complete file, but it may need revisions, voice rights, and stable pronunciation. That is why this comparison treats delivery path and operating constraints as first-class selection criteria.

What is a text-to-speech API?

A text-to-speech API turns written text into spoken audio through an HTTP or WebSocket endpoint, so an application generates voice programmatically instead of recording it. You send text and a voice ID, and the service returns audio, either as a complete file in batch mode or as a stream of chunks that start playing before generation finishes.

Modern TTS APIs do more than read text aloud. They produce natural prosody, handle pronunciation across languages, accept SSML or plain-language controls for pacing and emphasis, and stream audio incrementally to cut perceived latency.

For a voice agent, streaming is the capability that matters most because it starts playback on the first chunk rather than waiting for the full sentence, which is the difference between a reply that feels immediate and one that feels late.

Text-to-speech API evaluation methodology

Each API on this list is tested against five criteria:

1. Delivery path. I checked whether the API supports REST generation, streamed audio, WebSocket sessions, batch synthesis, or direct playback in a call flow.
2. Voice and language fit. I looked for documented voice inventory, locale coverage, pronunciation controls, styles, and custom voice options.
3. Integration surface. I compared SDKs, authentication, output formats, cloud dependencies, and the amount of application code required around synthesis.
4. Pricing transparency. I recorded published usage rates, free allowances, plan limits, and any model or region caveats. Where a vendor does not publish an API rate, we say so.
5. Production limits. I checked request size, concurrency, region availability, approval requirements, and streaming constraints that can change the implementation.

Text-to-speech APIs at a glance

1. Telnyx: one API across many voices, built for phone calls

The Telnyx Text-to-Speech API is built for speech in an app, phone call, or voice agent. It combines Telnyx and supported provider voices behind shared REST and WebSocket surfaces, then connects speech to Voice API and TeXML call flows.

Best for: real-time voice applications that need streaming, in-call playback, and provider choice.

Key features:

• REST synthesis with binary, base64, or asynchronous audio responses, including HTTP chunked audio
• Native telephony output: WebSocket audio in mulaw and alaw for 8kHz phone paths, alongside mp3, linear16, wav, and ogg_vorbis, with conversational barge-in built in
• Voice Design Lab for prompt-based custom voices and voice cloning, producing a voice ID usable across Telnyx Voice AI Agents, the Voice API for call control, and the TTS WebSocket API

Telnyx is a drop-in replacement for the OpenAI Audio API, so you can point an existing OpenAI TTS integration at it by swapping the base URL. Telnyx TTS will not top a voice-quality leaderboard, but it will power a production-grade phone stack for enterprise deployments.

Because the REST endpoint mirrors the OpenAI Audio API, an existing OpenAI TTS integration moves over by changing two lines:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_TELNYX_API_KEY",
    base_url="https://api.telnyx.com/v2"
)

response = client.audio.speech.create(
    model="tts-1-hd",
    voice="astra",
    input="Hello from Telnyx."
)

response.stream_to_file("output.mp3")

Limits:

• Dedicated voice models like ElevenLabs and Cartesia lead on raw expressive quality,
• The voice catalog spans providers, so quality is only as good as the model you route to

Pricing: Telnyx-owned voices run $0.000003 per character ($3/M) for Natural, $0.000048 for HD, and $0.000032 for Ultra. Routed third-party voices are billed per character through the same account, and they often cost less than going direct: Inworld Max is $0.00001/char ($10/M) through Telnyx versus $17.50 to $35/M on Inworld's own tiers, and Rime is $0.00002/char ($20/M) versus roughly $30 to $50/M direct (see the Telnyx TTS pricing page, for the entire breakdown).

Bottom line: choose Telnyx when the deciding factors are telephony delivery, compliance, and one integration across many voices.

2. Cartesia: the lowest time-to-first-audio on the list

Cartesia is a hosted streaming TTS and STT platform, aimed at engineering teams where response times are critical. Its Sonic model line, currently Sonic 3.5, runs on a State Space Model architecture instead of a transformer, which is how it holds latency down as load rises.

Cartesia pairs the API with Line, a managed voice-agent runtime for orchestration and deployment. That combination makes it the first option to test when the lowest possible time-to-first-audio is the deciding factor and your target languages stay within its coverage.

Best for: real-time phone and voice-agent stacks optimizing for the lowest possible time-to-first-audio.

Key features:

• Sonic 3.5 documents sub-90ms time-to-first-audio, and third-party comparisons peg Sonic's streaming TTFB as the fastest tested
• REST Bytes, SSE, and WebSocket streaming, with pcm_mulaw and pcm_alaw at 8000 Hz for telephony and sample rates up to 48kHz, plus Python and TypeScript SDKs
• Sonic 3.5 covers 42 languages; instant voice cloning trains from a clip up to 10 seconds on the free tier

Limits:

• 42 languages against ElevenLabs' 70-plus is a real gap for global deployments
• No documented on-prem deployment for the Line agent runtime, which is managed cloud only

Pricing: Free at $0/month (20K credits), Pro at $5/month (100K credits, adds commercial license and instant cloning), Startup $49/month, Scale $299/month. The Line voice-agent runtime bills at $0.06/min plus $0.014/min telephony.

Bottom line: choose Cartesia when raw response speed is the deciding factor and your languages stay inside its 42. If you need broader language reach or a documented on-prem path, look at Inworld or Rime.

3. Inworld: benchmarked quality at a low per-character rate

Inworld TTS is a hosted API from Inworld, which moved from character and NPC AI to a broader AI runtime and now runs its own TTS model family: Realtime TTS-2 plus TTS 1.5 Max and TTS 1.5 Mini. It is built for voice-first products that stream synthesis into a live conversation, with instant cloning and low documented latency.

Best for: real-time voice agents and conversational apps that need low-latency streaming TTS with instant cloning.

Key features:

• Three delivery paths: REST sync, HTTP streaming with progressive NDJSON chunks, and WebSocket for incremental LLM output, with MP3, WAV, or streaming-chunk output
• Documented streaming latency around 100 to 120ms median on TTS 1.5 Mini and about 200ms median on TTS 1.5 Max
• TTS-2 flagship documents 200+ languages and locales, natural-language voice steering, and instant cloning from 5 to 15 seconds of audio

Limits:

• No total prebuilt-voice count is publicly documented; the docs emphasize cloning and custom voices over a fixed catalog
• TTS-2, the most capable model, is a research preview rather than generally available, so the 15-language 1.5 models are the stable production path

Pricing: Priced per million characters, where roughly 1,000 characters is about a minute of audio. TTS-2 runs $25 per million on the free On-Demand tier down to $5 per million on Enterprise; TTS 1.5 Max runs $35 to $17.50; TTS 1.5 Mini runs $15 to $7. The On-Demand tier is free with 70 minutes included.

Bottom line: choose Inworld when you want benchmark-grade quality at consumer-app economics. If your deployment must run beyond its 15 production languages today, Fish Audio or ElevenLabs covers more ground.

4. Rime: the low-cost telephony voice you can self-host

Rime is a hosted TTS API purpose-built for conversational AI and phone voice agents, founded by linguists and marketed to contact centers across fintech, healthcare, and hospitality. Its models, currently Mist v3, Coda, and Arcana v3, target very low latency and telephony-native output, and it offers self-host and VPC deployment as first-class options. That combination makes it a candidate when the voice has to land on a phone call under a tight latency budget and compliance constraints.

Best for: phone and contact-center voice agents that need very low time-to-first-audio, native telephony output, and a self-host path.

Key features:

• REST synthesis, HTTP streaming, and WebSocket streaming with bearer-token auth, plus Python and JavaScript examples
• Telephony-ready output in mp3, mulaw, and pcm with sampling configurable from 4000 to 44100 Hz, so 8kHz mulaw for phone is supported
• Documented time-to-first-audio at single concurrency of 37ms on Mist v3 and 96ms on Coda, with self-host via Docker Compose or Kubernetes and VPC or on-prem for HIPAA and SOC 2

Limits:

• The pricing page is internally inconsistent, listing $0.03 per 1,000 characters in one place and $0.05 in another
• Batch synthesis and explicit on-prem pricing are not publicly documented

Pricing: Usage-based, listed between $0.03 and $0.05 per 1,000 characters depending on where the pricing page is read, with 3,000 free minutes for new accounts and 20 concurrent generations on the Starter plan. Enterprise is quote-only. Rime is also available through Telnyx at $0.00002/char ($20/M) with telephony attached.

Bottom line: choose Rime when cost-per-minute and telephony fit outweigh expressive range, and you want the option to self-host. If you need a voice that scores on naturalness, Inworld or Cartesia is the better trade.

5. Deepgram Aura: consistent voices that survive long calls

Deepgram is best known for speech-to-text, and Aura is its text-to-speech line, with Aura-2 as the enterprise real-time model. It is tuned for voice agents rather than expressive narration, prioritizing clarity and consistency across many conversational turns. Deepgram documents native telephony formats and an on-prem path, which fits agent builders who already think in terms of call audio and deployment control.

Best for: real-time voice agents that need low-latency, consistent synthesized speech with native telephony output and a self-host option.

Key features:

• REST synthesis that streams audio from the first byte, plus a WebSocket streaming interface, deployable in public cloud, private cloud, and on-premises
• Aura-2 documents 88 voices across 7 languages, with telephony encodings including mulaw and alaw that default to 8000 Hz
• SDKs for JavaScript, Python, Go, C#/.NET, and Java, with token-header authentication

Limits:

• A 2,000-character maximum per request means long text must be chunked by the caller
• The voice range is deliberately narrow and agent-oriented, not built for expressive multi-style narration

Pricing: Aura-2 costs $0.030 per 1,000 characters pay-as-you-go and Aura-1 costs $0.0150 per 1,000, with a $200 credit for new accounts. Self-hosted and enterprise rates are quote-only (Deepgram pricing page, July 2026).

Bottom line: choose Deepgram when you already run its STT or want one provider for the whole speech loop with an on-prem option. If you need expressive range or multilingual reach beyond 7 languages, ElevenLabs or Fish Audio covers more.

6. Fish Audio: broad-language cloning with an open-weight fallback

Fish Audio is a hosted voice platform that exposes text-to-speech, speech-to-text, and voice cloning through REST and WebSocket surfaces, alongside an open-weight model line called OpenAudio. Its distinction is cost and reach: broad language coverage and fast cloning at a low per-byte rate, with the option to fall back to an open model for self-hosting.

That makes it a fit for teams optimizing spend across high synthesis volume, though commercial use of the open model routes through the hosted API.

Best for: teams that want low-cost, broad-language hosted TTS with fast voice cloning and an optional open-weight fallback.

Key features:

• REST synthesis with chunked HTTP streaming and WebSocket streaming with low, balanced, and normal latency modes, plus Python and JavaScript SDKs
• Flagship s2.1-pro documents 83 languages, s2-pro documents 100ms time-to-first-audio, and the public library advertises 2 million community voice models
• Zero-shot voice cloning from 10 to 30 seconds of reference audio

Limits:

• Concurrency is tied to account spend: 5 concurrent requests under $100 paid, 15 at $100 or more, and 50 at $1,000 or more
• The open-weight OpenAudio S1-mini model is non-commercial under CC-BY-NC-SA-4.0, so commercial use runs through the hosted API

Pricing: All hosted TTS models cost $15.00 per 1 million UTF-8 bytes, roughly 180,000 English words or about 12 hours of speech, on a pay-as-you-go basis. A free s2.1-pro tier runs under fair-use limits.

Bottom line: choose Fish Audio for cheap, broad-language hosted synthesis with an open fallback for research. Look elsewhere if you need commercial use of the open model.

7. ElevenLabs: the voice-realism benchmark

ElevenLabs provides a hosted TTS API through HTTP and WebSocket requests, with maintained Python and Node.js libraries. Its main distinction is voice identity: the current developer materials combine a large voice catalog with cloning and expressive model options. That focus makes it attractive when the voice itself is part of the product.

It does not answer telephony, call control, or the rest of a live voice stack, so teams must own that boundary.

Best for: teams where a recognizable voice, cloning workflow, or expressive delivery is the primary buying criterion.

Key features:

• HTTP and WebSocket APIs with official Python and Node.js libraries
• A catalog of 10,000+ voices and voice-cloning workflows in the developer quickstart
• Flash/Turbo and Multilingual model families, with 32 listed languages for those models

Limits:

• The current API pricing page lists a 40,000-character request limit for Flash/Turbo and Multilingual v2/v3 models
• Voice rights, cloning consent, retention, and model access need a separate commercial review before launch

Pricing: Pay-as-you-go pricing is $0.05 per 1,000 characters for Flash/Turbo and $0.10 per 1,000 characters for Multilingual v2/v3.

Bottom line: choose ElevenLabs when voice realism outweighs owning telephony and a single bill. If per-character cost predictability or a carrier-native path matters more, Inworld or Telnyx fit better.

8. Google Cloud Text-to-Speech: GCP-native, batch-first

Google Cloud Text-to-Speech is a managed synthesis API for teams already operating inside Google Cloud. It accepts text or SSML through REST, gRPC, and client libraries, with pitch, speaking-rate, volume, and format controls. Its main advantage is cloud fit: identity, billing, monitoring, and deployment can stay in the same environment as the rest of the application. Live-use teams still need to validate the selected model's streaming path.

Best for: Google Cloud applications that need a broad catalog of documented voices and language variants.

Key features:

• 380+ voices across 75+ languages and variants
• SSML, pitch, speaking-rate, volume, and audio-format controls for structured speech
• REST, gRPC, and Google Cloud client libraries for common application languages

Limits:

• Google documents bidirectional streaming as preview and limits it to Chirp 3: HD voices, so it is not a universal replacement for synchronous synthesis
• The pricing catalog spans several model families, and Google Cloud infrastructure costs sit outside the TTS line item

Pricing: Standard costs $4 per 1 million characters after 4 million free. Neural2 costs $16 per 1 million after 1 million free; newer models use different rates.

Bottom line: choose Google Cloud when GCP integration and language breadth outweigh real-time performance. For a live phone agent, a dedicated streaming provider on this list will serve the call path better.

Which TTS API should you choose?

The five criteria that matter at scale for a production-ready TTS API:

• Latency (time-to-first-audio). For a live agent, response time above roughly 300ms end-to-end starts to feel like a delay. Check the published TTFA, then measure it in your own pipeline, because synthesis-only numbers hide the network and telephony hops.
• Telephony output. A phone path needs 8kHz mulaw or alaw. A provider that only returns studio-rate mp3 forces a re-encode on every turn, which adds latency and cost.
• Language and voice coverage. Count the languages you actually ship in, not the headline total, and confirm the specific accents and mixed-language behavior your callers will hear.
• Cost at real volume. Normalize every quote to one million characters, then add telephony, STT, and LLM. A low character rate can still produce a higher bill once the full pipeline runs.
• Deployment and compliance. If you need HIPAA, data residency, or on-prem, confirm it is documented and available.

With these in mind, here's how each tool stacks up:

• Choose Telnyx if speech has to reach a caller through a real phone path, you want one API across many voices, and compliance is a key priority.
• Choose Cartesia if time-to-first-audio is the deciding number.
• Choose Inworld if you want top-10 benchmarked quality at a low per-character rate and can start on its 15 production languages.
• Choose Rime if cost-per-minute and telephony fit matter more than expressive range, and you want a self-host option.
• Choose Deepgram if you want one provider for STT and TTS with an on-prem path and consistency over flair.
• Choose Fish Audio if you need cheap, broad-language cloning and value an open-weight fallback for research.
• Choose ElevenLabs if voice realism and cloning are the product and you will own telephony yourself.
• Choose Google Cloud if you are GCP-native and the work is batch narration rather than a live call.

Match the API to the industry workflow

• Contact centers and voice agents: prioritize streaming, interruption handling, call-compatible formats, regional processing, and an escalation path when a provider fails.
• E-learning and training: prioritize consistent voice identity, long-form generation, pronunciation, revision speed, and commercial rights across a library of lessons.
• Content creation: prioritize expressive controls, cloning consent, export formats, editing workflow, and predictable cost for repeated drafts.
• Healthcare and finance: start with data handling, region, retention, access controls, contracts, and documented compliance.

Test before launch

1. Send production-shaped text with names, addresses, dates, numbers, URLs, abbreviations, and domain terms.
2. Measure time to first playable audio and total turn time with the exact voice, region, format, and interface you plan to use.
3. Interrupt a response, send a new turn, and check whether buffered audio stops without leaking into the next response.
4. Repeat the test in every target locale and record pronunciation, accent, mixed-language, and number behavior.
5. Run at expected concurrency and capture quotas, rate limits, retries, connection failures, and fallback behavior.
6. Calculate the complete cost, including TTS characters, telephony, STT, model inference, orchestration, storage, transfer, and provider-key charges.

Independent leaderboard ranking

The most-cited neutral quality reference for TTS APIs is the Artificial Analysis Speech Arena, a blind human-preference ELO board.

As of July 31, 2026 it placed Google's Gemini 3.1 Flash TTS at #3 (Elo 1,212), Cartesia Sonic 3.5 at #5 (1,202), Inworld Realtime TTS 1.5 Max at #7 (1,196) and TTS-2 at #9 (1,189), ElevenLabs Eleven v3 at #10 (1,173), and Fish Audio's best model at #14 (1,139). Rime's Coda sat at #53 (1,046).

Two vendors on this list, Telnyx and Deepgram Aura, do not appear on the leaderboard. That is not an oversight to hide; it reflects what those two optimize for: production reliability and telephony fit, rather than blind-preference voice quality.

If a benchmark ELO is your only criterion, the leaderboard leaders are the honest answer. For most production voice agents, it is one input among latency, telephony, and cost.

Where the big cloud providers fit

Amazon Polly and Azure Neural voices are not ranked separately here because they behave as components rather than destinations. Both are available through the Telnyx API (Polly hosted, Azure as bring-your-own-key), and both are reasonable general-purpose cloud TTS when you do not need agent-grade latency or telephony output.

Frequently asked questions

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

The Telnyx TTS API is the best fit for voice agents that need to make and receive live phone calls. Deepgram is also purpose-built for this use case, and offers on-prem deployment. Cartesia and Rime post the lowest published time-to-first-audio (sub-90ms and 37ms) and can also be part of your consideration stack.

What is the cheapest TTS API for voice agents?

On raw per-character rate, Inworld Mini through Telnyx is among the lowest at $0.0000055 per character ($5.50/M), and Rime and Deepgram Aura sit around $0.02 to $0.03 per 1,000 characters. However, the cheapest headline rate is not always the cheapest bill since telephony, STT, and LLM costs add up.

Which TTS API has the best voice quality?

On the independent Artificial Analysis Speech Arena as of July 2026, the top of the board included Google's Gemini 3.1 Flash TTS, Cartesia Sonic 3.5, and Inworld's Realtime models, with ElevenLabs Eleven v3 close behind. The board reshuffles often, so verify the current standing and run your own blind test on production-shaped text before deciding.

Do I need HIPAA compliance for my voice agent?

Yes, HIPAA compliance is mandatory if your voice agent touches protected health information in the United States, such as appointment scheduling, insurance verification, or patient communication. Telnyx, Deepgram, Rime, and Inworld all document HIPAA-eligible paths, though several gate BAAs behind enterprise or add-on tiers.

What is the difference between a TTS API and a voice API?

A TTS API turns text into audio. A voice API adds the communications layer around that audio: call control, phone numbers, SIP, media streaming, and in-call playback. Use a TTS API when your application already owns audio delivery. Use a voice API when speech has to reach a caller as part of a programmable call flow, which is why Telnyx pairs the two behind one account.

How do I get telephony-grade audio out of a TTS API?

You need 8kHz mulaw or alaw output that a phone network can carry without re-encoding. Telnyx, Cartesia, Rime, and Deepgram all document native mulaw output; Fish Audio supports 8kHz PCM. Providers that only return studio-rate mp3 or PCM force a conversion step that adds latency on every turn.

Are there free text-to-speech APIs?

Most providers on this list offer a free tier or credits for evaluation: Cartesia is free at $0/month (20K credits), Inworld includes 70 free minutes, Rime gives 3,000 free minutes, Deepgram grants a $200 credit, Fish Audio has a fair-use free tier, and Google Cloud includes 1M free characters a month. Telnyx has no free tier and uses volume discounts instead. Free tiers carry rate limits and no SLA, so use it for demo runs exclusively.

What is the difference between real-time and batch TTS?

Real-time (streaming) TTS delivers audio in chunks as it generates, so playback starts on the first chunk, which is what a voice agent needs to feel responsive. Batch TTS processes the whole input and returns a complete file, which suits audiobooks, narration, and pre-recorded content where a second of startup delay does not matter. Most APIs here support both; confirm the streaming path on the specific model you plan to ship.

What is the fastest text-to-speech API?

On published time-to-first-audio, Rime leads at 37ms on its Mist v3 model at single concurrency, with Cartesia's Sonic 3.5 at sub-90ms and Deepgram Aura near 90ms after optimization. Treat these as synthesis-layer figures, not end-to-end: the network hop and telephony leg add latency on every turn, so measure the full path in your own pipeline before committing.

Which TTS API handles accents and dialects best?

Match the accent to a provider's documented coverage, not its headline language total. Deepgram Aura-2 ships American, British, Australian, Irish, and Filipino English plus Spanish voices that code-switch; Google Cloud spans 75+ languages and regional variants; ElevenLabs covers roughly 74 languages and Fish Audio 83. Test the exact accents and mixed-language lines your callers use, since a high language count rarely predicts quality on one specific dialect.

Which TTS API is best for customer support chatbots?

Customer support chatbots need streaming latency, telephony output when calls are involved, and clean SDK integration with your existing speech-to-text and LLM workflows. Cartesia, Rime, and Deepgram are built for that real-time loop, and Telnyx adds telephony plus one bill across STT, TTS, and call control. For chat-only support where no phone leg exists, latency matters less and voice quality or cost can lead the decision.

Which TTS APIs work for custom web applications?

REST APIs work well for applications that request a file or audio response. Streaming APIs matter when the browser or app should begin playback before synthesis finishes. Check authentication, CORS or server-side proxy requirements, audio formats, caching, quota behavior, and whether the selected voice supports the interface.