Best Text-to-Speech APIs for Developers (2026): Latency, Pricing, and Cloning

This guide is for choosing a text-to-speech API to build on, not a studio to sit down and record in. The needs are different: you care about latency under load, cost per million characters, whether cloning is self-serve or a sales call, and whether you can run the thing inside your own network.

If you want an editor that produces finished voiceover files, read Best AI Text-to-Speech Software (2026) instead. That guide covers the creator and consumer apps. This one covers the metered endpoints behind them.

Pricing in this space moves constantly, so treat every figure here as a starting point and confirm current rates on the provider's own pricing page before you commit to an architecture.

How We Picked

Four dimensions decide most API choices, in roughly this order:

1. Deployment model. Hosted cloud endpoint, self-hosted, or on-prem. This is usually a hard constraint set by your compliance or privacy requirements, so it filters the list before anything else does.
2. Latency. Time to first audio is what makes a live conversation feel natural or broken. It matters enormously for agents and phone bots, and not at all for a nightly batch job.
3. Cost at volume. Per-million-character pricing diverges by nearly an order of magnitude across this list, and the gap compounds fast once you are generating real volume.
4. Cloning access. Whether you can clone a voice programmatically on a self-serve plan, or whether it is gated behind an enterprise agreement and a consent process.

Quick Comparison

API	Deployment	Latency	Cloning	Languages	Price	Best for
Voice Creator Pro	Local, on your hardware	Depends on your GPU	Yes, instant, every tier	600+	$54.99 one-time, unmetered	Offline and on-prem pipelines
Cartesia (Sonic-3)	Hosted	Very low (~40ms TTFA)	Yes, from Pro	40+	Free, then $5/mo	Realtime voice agents
Google Cloud	Hosted	Moderate	Yes, instant, consent-gated	75+	~$4 to $30/1M chars	Cheapest multilingual batch
Microsoft Azure	Hosted	Moderate	Enterprise only	140+	~$16/1M chars	Widest language coverage
Amazon Polly	Hosted	Moderate	No self-serve	~40 locales	$4 to $30/1M chars	High-volume AWS-native work
OpenAI	Hosted	Low to moderate	None	50+	~$0.015/min	Promptable delivery, prototypes
Deepgram Aura-2	Hosted + on-prem	Low	No, fixed voices	English-focused	~$30/1M chars	Regulated and on-prem agents
ElevenLabs	Hosted	Moderate	Yes, from Starter	70+	Free, then $6/mo	Maximum realism
Resemble AI	Hosted	Moderate	Yes, core product	40+	Pay-as-you-go	Bespoke cloned brand voices
Rime / LMNT	Hosted	Low	LMNT self-serve, Rime enterprise	LMNT 30+, Rime English	~$35/1M (LMNT)	Budget voice agents

1. Voice Creator Pro

Best for: offline, on-prem, and privacy-constrained pipelines where you do not want audio leaving your network or a meter running.

Be clear on what this is, because it is a different shape from everything else on this list. Voice Creator Pro Desktop ships a local REST API: you run the app on your own Windows or macOS machine and call an endpoint on that machine. There is no hosted cloud endpoint to point a production web service at, and no autoscaling. What you get instead is an unmetered TTS endpoint that works with no internet connection.

• Deployment: local only. The desktop app is the server. Nothing is sent to a third party.
• Pricing: one-time $54.99 to $59.99 for the desktop app, with unlimited generations. No per-character billing and no monthly cap.
• Cloning: zero-shot from a 3 to 10 second reference clip, self-serve, no consent workflow, no training step. Available on every tier including the free Cloud tier. Longer reference audio does not improve the result.
• Languages: 600+ for cloning and voice design, the widest coverage on this list.
• Emotion: 13 selectable emotions with intensity control, plus prompt-based delivery direction on models like DramaBox and Qwen3-TTS.

Where it fits: build pipelines, game engines, batch generation, internal tooling, and anything handling confidential scripts. If your constraint is "this audio cannot leave our infrastructure" or "we cannot have a per-character bill", this solves it in a way the hosted APIs structurally cannot.

Honest cons:

• It is not a hosted cloud API. You cannot point a serverless function at it without running and exposing your own machine.
• No autoscaling, no SLA, no multi-region. Throughput is whatever your hardware does.
• Not built for sub-100ms realtime agent latency. For a live phone bot, use Cartesia or Deepgram.

See the Voice Creator Pro API documentation for endpoint details, or try the models free in your browser before buying the desktop app.

2. ElevenLabs

Best for: the highest voice realism available from an API, with mature SDKs.

ElevenLabs is the realism benchmark, and its API is well documented with real SDK support. If the quality of a single generated line is what your product is judged on, this is the default answer.

• Pricing: Free ($0, about 10 minutes a month), Starter ($6/mo, ~30 min), Creator ($22/mo, ~2 hrs), Pro ($99/mo, ~10 hrs), Scale ($299/mo, ~30 hrs), Business ($990/mo, ~100 hrs). Overage runs roughly $0.17 to $0.36 per minute by tier.
• Cloning: instant cloning from Starter, higher-fidelity professional cloning from Creator.
• Languages: up to 70+ depending on model.
• Commercial rights: included from Starter up.

Watch for: the audio allowance disappears quickly at production volume, and the effective cost per minute is far above the raw cloud APIs. Quality can also wobble on very long passages, so chunk long text. For a full head-to-head, see Voice Creator Pro vs ElevenLabs.

3. Cartesia (Sonic-3)

Best for: realtime voice agents, IVR, and phone bots where latency is the product.

Cartesia is purpose-built for live conversation. Its Sonic-3 model reports time-to-first-audio around 40ms, which is the difference between an agent that feels responsive and one that feels broken.

• Pricing: Free ($0, 20,000 TTS credits a month, no cloning), Pro ($5/mo, 100,000 credits or roughly 133 minutes, instant cloning plus commercial license), Startup ($49/mo, ~1,667 min, higher-fidelity cloning), Scale ($299/mo, ~10,667 min), Enterprise (custom). Live voice-agent calls bill separately at about $0.06 per minute.
• Cloning: instant from Pro, higher fidelity from Startup.
• Languages: around 40.
• Extras: streaming output, emotion and laughter cues, production-oriented SDKs.

Watch for: the separate per-minute billing on live agent calls is easy to miss when modelling costs. It is also the wrong tool for offline narration work, where you are paying a latency premium you will never use.

4. Google Cloud Text-to-Speech

Best for: the cheapest multilingual speech at volume.

• Pricing: pay-as-you-go, roughly $4 per million characters for standard and WaveNet voices, up to about $30 per million for Chirp 3 HD, with a monthly free allowance. That works out to roughly 35 to 250 minutes per dollar depending on voice tier.
• Cloning: Chirp 3 Instant Custom Voice clones from about 10 seconds of audio, consent-gated, with a heavier enterprise option above it.
• Languages: 75+ with SSML support.

Watch for: per-voice quality is uneven across the catalogue, and emotional steering is noticeably flatter than ElevenLabs or Hume. Audition the specific voices and locales you need rather than trusting the headline count.

5. Microsoft Azure TTS

Best for: localization at scale and precise, tag-level prosody control.

• Pricing: about $16 per million characters (neural), around $22 per million for HD, commitment tiers down to about $7.50 per million, plus 0.5M characters a month free.
• Cloning: custom neural voice exists but is an enterprise, consent-gated engagement, not self-serve.
• Languages: 140+, the widest of any hosted API here.

Watch for: if you need programmatic cloning on a self-serve plan, Azure will not give it to you. Its strength is breadth and SSML precision, not fast cloning.

6. Amazon Polly

Best for: high-throughput app narration, notifications, and IVR inside AWS.

• Pricing: $4 per million characters (standard), $16 per million (neural), $30 per million (generative), with 1M neural characters a month free for the first year.
• Cloning: none self-serve. Brand Voice is a custom AWS engagement.
• Languages: around 40 locales, with SSML for pacing and pronunciation.

Watch for: it will not fool anyone into thinking it is human, and emotional range is minimal. That is a fair trade when you are generating notification audio at scale inside an AWS stack.

7. OpenAI TTS

Best for: promptable delivery and fast prototypes.

• Pricing: about $0.015 per minute on gpt-4o-mini-tts. The older tts-1 is $15 per million characters, tts-1-hd $30 per million. No subscription.
• Cloning: none. Fixed voice set.
• Languages: 50+.

Watch for: you steer delivery with a prompt ("calm support agent", "excited narrator") rather than SSML, which is either the appeal or the limitation depending on how much control you need. Input length per request is short, so long text needs chunking.

8. Deepgram Aura-2

Best for: regulated and on-prem agent deployments.

• Pricing: $0.030 per 1,000 characters (about $30 per million), pay-as-you-go.
• Cloning: none. Fixed preset voices.
• Languages: English-focused, with multilingual support still limited.

Watch for: the on-prem option is the real differentiator here. If you are not compliance-constrained, you can get comparable agent latency more cheaply elsewhere.

9. Resemble AI

Best for: building bespoke cloned brand voices into your own product.

Resemble is cloning-first, with real-time speech effects and deepfake detection layered on top. The team also maintains the open-source Chatterbox and DramaBox models.

• Pricing: Flex pay-as-you-go from $0 (about $0.0005 per second, voice clones $2 to $5 a month each), or Enterprise (custom). The older Creator and Professional subscription tiers were discontinued in 2025.
• Cloning: yes, this is the core of the product.
• Languages: 40+.

Watch for: the flow is enterprise-oriented, so it can feel heavy if you just want to generate audio. There is no free tier. See our full Resemble AI comparison.

10. Rime and LMNT

Best for: budget voice agents where speed and cost beat library size.

• Pricing: LMNT is about $35 per million characters. Rime sells per-use through its API. Both offer free playgrounds.
• Cloning: LMNT offers self-serve instant cloning from a short sample. Rime handles cloning as an enterprise custom voice offering.
• Languages: LMNT covers 30+, Rime is English-focused with a large prebuilt speaker library.

Watch for: smaller voice catalogues and less mature tooling than the majors. Fine for agents, thin for anything that needs voice variety.

What Latency Actually Means Here

Latency gets quoted as a single number, but there are three that matter and they are not interchangeable.

Time to first audio (TTFA) is how long until the first byte of speech comes back. This is the one that governs whether a conversation feels natural. Under roughly 200ms and a human will not notice; past about 500ms the pause reads as the system being confused. Cartesia's ~40ms is the headline figure on this list.

Streaming versus complete generation. A streaming API returns audio as it generates so playback starts immediately. A non-streaming API returns a finished file, so your user waits for the whole thing. For a 30-second narration, a non-streaming API can be several seconds of dead air. Any API you use for live conversation must stream.

Throughput under concurrency. A benchmark run against an idle endpoint tells you very little about behaviour at 200 concurrent sessions. Rate limits and queuing usually bite before raw model speed does, so load-test your actual concurrency before committing.

For batch work, all of this is irrelevant. If you are generating audiobook chapters overnight, optimise purely for cost per character and voice quality, and ignore latency entirely.

Cost at Scale: A Worked Example

Take a product generating 10 million characters a month, roughly 200 hours of speech. The spread is dramatic:

Provider	Approximate monthly cost
Google Cloud (standard/WaveNet)	~$40
Amazon Polly (neural)	~$160
Microsoft Azure (neural)	~$160
Deepgram Aura-2	~$300
LMNT	~$350
ElevenLabs	Several thousand at published per-minute rates
Voice Creator Pro Desktop	$54.99 to $59.99, one time

Two things follow. First, the realism premium is real and large: ElevenLabs is the best-sounding option here and costs one to two orders of magnitude more than Google Cloud at volume. Whether that is worth it depends entirely on whether your users are judging voice quality or just consuming information.

Second, if your workload can run on your own hardware, the economics change completely, because you are buying a machine rather than renting characters. That only works when you control the environment and do not need elastic scale, which rules it out for most consumer web products and makes it compelling for internal tooling, game build pipelines, and regulated environments.

How to Choose by Use Case

• Realtime voice agents and phone bots: Cartesia for the lowest latency, Deepgram when you need on-prem or regulated deployment, Rime or LMNT on a tight budget.
• High-volume batch generation: Google Cloud is the cheapest per character, Amazon Polly if you are already AWS-native.
• Localization across many languages: Azure (140+) or Google Cloud (75+) for hosted, Voice Creator Pro (600+) if you can run locally.
• Maximum voice realism: ElevenLabs, with the cost consequences above.
• Programmatic voice cloning: Voice Creator Pro (every tier, no consent step), ElevenLabs (from Starter), Cartesia (from Pro), Resemble AI (core product), or LMNT. Azure and Amazon gate cloning behind enterprise agreements.
• Offline, air-gapped, or confidential: Voice Creator Pro Desktop's local REST API, or Deepgram's on-prem deployment. Everything else on this list sends your text to a third party.
• Prototyping: OpenAI gpt-4o-mini-tts is the fastest thing to wire up, and cheap enough to leave running.

---

Try Voice Creator Pro free in your browser, or get the desktop app for unlimited offline generation with a local REST API. Both include full commercial rights.