Best Text-to-Speech APIs for AI Apps

As of July 28, 2026, ElevenLabs is our default shortlist when an AI product needs streaming text-to-speech, voice cloning, and an agent platform from one vendor. Cartesia deserves a separate trial when the team wants more control over a latency-focused streaming path. Google Cloud Text-to-Speech and Amazon Polly make sense when the application already lives inside those clouds.

That is a product-fit verdict, not a voice-quality benchmark. We have not run the same scripts, voices, regions, and playback clients across every API. We will not turn vendor demos into a score table.

The product evidence below comes from current first-party documentation. Use the ElevenLabs pricing calculator to turn character volume and peak concurrency into a plan estimate.

This roundup covers the speech layer of the AI App Stack. If you are choosing the model and orchestration path, start with Best LLM for Voice Agents. A fast voice cannot rescue a slow or unreliable conversation loop.

We rank product fit, not voices

• Streaming path: HTTP chunks, WebSockets, connection reuse, and controls that affect when audio starts.
• Voice control: stock voices, cloning, pronunciation, pacing, and model versioning.
• Operational fit: regional processing, concurrency, observability, and how much of the agent stack one vendor owns.
• Commercial shape: usage billing, free-plan access, and whether commercial rights differ by plan.
• Exit cost: how much voice identity, code, and evaluation work must be repeated after a switch.

Naturalness belongs in the criteria, but it needs a listening panel and the same script. Until we run one, it stays a test to perform rather than a number to publish.

The table is documented surface, not a listening panel

Table 1 Download CSV

Tool	Best for	Billing shape	Prototype path	Verified standout
ElevenLabs	One vendor for TTS, cloning, and agents	Usage within plan tiers	Published free plan; check current rights and quota	HTTP/WebSocket streaming, Flash and multilingual models, two cloning paths
Cartesia	Teams tuning a realtime streaming path	Usage plans	Check live console and pricing	Multiplexed WebSocket contexts and versioned Sonic models
OpenAI Realtime	OpenAI-native speech-to-speech applications	Token and audio usage	API billing	Native realtime audio path
Google Cloud TTS	Google Cloud estates and conventional synthesis	Character usage by voice class	Cloud free-usage terms	Cloud IAM, regions, and service integration
Amazon Polly	AWS estates, IVR, and read-aloud features	Character usage by engine	AWS free-tier terms	AWS integration and speech marks
Kokoro	Teams willing to self-host an open-weight model	Infrastructure cost	Run it yourself	Model weights and local control

Exact quotas and prices rot quickly. Follow each vendor's live pricing page before forecasting production cost.

ElevenLabs is the full-stack shortlist

ElevenLabs documents streaming over HTTP and WebSockets, Flash models for lower inference latency, multilingual speech models, instant and professional voice cloning, and a separate ElevenAgents platform. That breadth is why it starts our shortlist: a team can test a TTS endpoint, then decide whether to keep composing the stack or use the same vendor for turn-taking, tools, telephony, and monitoring.

The published latency material is unusually clear about its own boundary. ElevenLabs reports roughly 75 ms for Flash model inference, while its latency documentation says actual time-to-first-byte depends on region, endpoint, voice, networking, and load. We treat 75 ms as a vendor model figure, not a promise about a user's first audio.

Where it loses: consolidating speech and agents with one vendor raises switching cost, professional clones can add latency, and commercial use depends on the plan. If the application needs only plain notification speech, the extra surface may solve problems it does not have.

Cartesia is the latency-engineering shortlist

Cartesia's current WebSocket documentation recommends opening the connection before the first generation, using separate contexts for turns, and starting a new context after an interruption. That is useful operational guidance, not merely a fast-demo claim.

The 2026 changelog also warns that sonic-3-latest can change without notice and recommends a dated snapshot for production. That earns trust because it names the exact trade-off: preview aliases improve quickly; production pins should not.

Where it loses: the streaming controls create work. Buffering too little can damage prosody, while buffering too much delays the first audio. A team that wants a managed agent rather than a TTS component may prefer a broader platform.

Cloud TTS wins when the cloud is the constraint

Google Cloud Text-to-Speech and Amazon Polly belong on the list because identity, billing, regional architecture, and support contracts often outweigh a marginal voice preference. If the feature reads alerts, accessibility text, or fixed prompts, using the cloud already approved by security can be the shortest route to production.

We have not run a current naturalness or cost benchmark across their voice classes. So we are not claiming either is cheaper or less natural in the abstract. Price the exact engine, region, output, and volume on the official calculator.

OpenAI Realtime is a different shape: speech and model reasoning share a native realtime API. It can remove integration boundaries, but it also narrows component choice. Compare it at the platform level, not as though it were only a drop-in TTS endpoint.

Run the test vendors cannot run for you

Use 20 to 50 scripts drawn from the product:

Table 2 Download CSV

Script class	What to listen for
Names, addresses, order codes	Pronunciation and correction behavior
Interrupted sentence	Cutoff, resume, and duplicated words
Long support answer	Drift, pacing, and consistency
Numbers, URLs, and units	Text normalization
Two target languages	Accent and code-switching

Measure application request to first audible audio, not the vendor's inference statistic. Keep P50 and P95, and run from the regions where users live. Record retries, malformed audio, and the human preference vote. A tiny reproducible test beats a sweeping claim about realism.

The pick is the stack you already have

• TTS, cloning, and an optional agent platform in one account: ElevenLabs, after checking live plan rights.
• A team prepared to tune WebSocket buffering and pin model snapshots: Cartesia.
• A native OpenAI speech-and-reasoning path: OpenAI Realtime.
• An existing Google Cloud or AWS operating boundary: trial that cloud's TTS before adding another vendor.
• Local control and an engineering team willing to operate inference: Kokoro or another open-weight model, after a license and hardware review.