Text-to-speech quietly became one of the more competitive corners of the AI API market. Quality has gone up, latency has come way down for the models built for real-time use, and pricing spans a genuinely wide range depending on what you're optimizing for. This roundup compares the major text-to-speech APIs on the things that actually affect a build decision: price, voice cloning support, latency, and where each one fits best.
If you're also working with AI-generated video, music, or images, it's worth thinking about TTS as part of a larger media pipeline rather than as a standalone feature. Apiframe's AI Video Generation API guide explains how developers can build similar media-generation workflows through an API.
What to Look for in a Text-to-Speech API
Voice quality and naturalness. This is subjective and worth testing with your own scripts rather than trusting any provider's demo reel, since demos are cherry-picked by definition. Pay attention to how a voice handles pauses, emphasis, and less common words, not just clean, simple sentences.
Latency: real-time vs batch. If you're building a voice agent or anything conversational, time-to-first-audio matters a lot, and the gap between providers is large. If you're generating narration for a video or podcast ahead of time, batch latency barely matters.
Voice cloning support. Some providers offer instant cloning from a short sample, others require more setup or a verification step, and some don't offer it at all. This is also where consent and licensing questions become real, more on that below.
Language and accent coverage. Varies a lot by provider. If you need strong non-English support, check this specifically rather than assuming a provider that's great in English will be equally good elsewhere.
Licensing for commercial use. Read the terms before you build. Some providers restrict how cloned voices can be used commercially, and that restriction usually lives in a licensing page most people never check until it's a problem.
Pricing model. Per-character, per-minute, and subscription tiers all show up. They're not directly comparable without doing the conversion yourself, which the table below does for you.
Quick Comparison Table
| Provider | Starting price | Voice cloning | Latency | Best for |
|---|---|---|---|---|
| ElevenLabs | ~$50/1M chars (Flash), ~$100/1M chars (Multilingual v2) | Yes, instant | Low (Flash tier) | Voice cloning quality, multilingual |
| OpenAI TTS | ~$15/1M chars (standard), ~$30/1M chars (HD) | No | Moderate | Simplicity, OpenAI ecosystem fit |
| PlayHT | ~$30-40/1M chars | Yes | Moderate | High-volume narration, audiobooks |
| Cartesia | ~$5-37/1M chars depending on plan | Yes | Very low (~40ms) | Real-time conversational voice |
| Google Cloud TTS | Varies by model tier | Limited | Low | Language and accent breadth |
| Amazon Polly | ~$4/1M chars (standard voices) | No (native) | Low | Enterprise, AWS-native pricing |
| Deepgram Aura-2 | $30/1M chars (pay-as-you-go) | No | Low | Developer-first, API-only positioning |
Prices shift often enough that it's worth checking each provider's current pricing page before committing, but the relative ordering here is a reasonable starting point.
The Best Text-to-Speech APIs
ElevenLabs
ElevenLabs is generally considered the voice cloning leader, with instant cloning from a short sample and consistently high naturalness scores across independent comparisons. It's priced at the higher end of the market (Multilingual v2 runs about $0.10 per 1,000 characters, more than triple Deepgram's rate), but the Flash v2.5 model cuts that close to $0.05 per 1,000 characters if you don't need the top-tier model for every request. Apiframe's ElevenLabs API guide covers pricing and setup in more depth if you want the full breakdown, and Apiframe also offers ElevenLabs Music generation through the same unified endpoint used for its other audio and video models.
OpenAI TTS
If you're already building on OpenAI's ecosystem, OpenAI's TTS models are the path of least resistance: same API key, same SDKs, straightforward pricing at roughly $15 per million characters for the standard tier and $30 per million for HD quality. It doesn't lead on any single metric, voice cloning isn't supported, but the simplicity is the actual selling point here.
PlayHT
PlayHT positions itself around real-time streaming and a large voice library (900-plus voices), with pricing in the $30-40 per million character range on its higher-volume plans. It's a solid fit for high-volume narration work like audiobooks and long-form podcast content, where breadth of voice options matters more than shaving milliseconds off latency.
Cartesia
Cartesia's Sonic model is built specifically for low-latency, real-time use, with a time-to-first-audio figure quoted around 40 milliseconds, among the fastest in the category. Pricing is credit-based and comes out to roughly $5 to $37 per million characters depending on plan tier. If you're building a voice agent where the user is actively waiting on a response, this latency profile is the main reason to pick Cartesia over a slower but similarly priced alternative.
Google Cloud Text-to-Speech
Google's strength here is breadth: more languages and regional accents than most competitors, backed by Google's translation and language infrastructure. Pricing varies significantly depending on which underlying model you use (Google offers several voice quality tiers), so it's worth checking current rates for the specific voice tier you need rather than assuming a single number applies across the board.
Amazon Polly
Polly is the obvious choice if you're already deep in AWS, both for the native billing integration and for enterprise procurement reasons. Pricing is around $4 per million characters for standard voices, making it one of the cheaper options in this list, though it trails the newer entrants on naturalness and doesn't offer native voice cloning.
Deepgram Aura-2
Deepgram built Aura-2 with an API-first, developer-focused positioning rather than a consumer app layered on top. Pricing sits at $30 per million characters on pay-as-you-go, dropping slightly on higher-volume plans. It's a reasonable middle ground on price and latency if you don't need voice cloning and want a straightforward, well-documented API.
Text-to-Speech API Pricing Models Explained
Per-character pricing (usually quoted per 1,000 or per million characters) is the most common model and the easiest to estimate costs from ahead of time, since your input text length is known before you make the call. Per-minute pricing is less common for pure TTS but shows up occasionally, and only really makes sense to compare once you know your average characters-per-minute of speech (roughly 800-1000 characters per minute of natural speech, though this varies by language and speaking rate). Subscription tiers bundle a fixed character allotment into a monthly price, which is usually cheaper per character than pay-as-you-go if your volume is predictable, but wasteful if it isn't.
To estimate monthly cost: take your expected monthly audio output in minutes, multiply by roughly 900 characters per minute as a rough average, then multiply by the provider's per-character rate. A podcast generating 10 hours of narration a month, for example, works out to around 540,000 characters, which lands well within most providers' starter or pay-as-you-go tiers. If you're already working with other AI media APIs, Apiframe's AI Music API guide explains a similar approach to comparing pricing, licensing, and integration requirements for AI-generated audio.
Voice Cloning: What to Check Before You Build
Voice cloning raises consent and licensing questions that are worth taking seriously before you ship a feature around it, not after. Reputable providers require some form of proof that the person being cloned has consented, and several now attach provenance metadata to generated audio so it's traceable back to its source. In the US, the FCC has classified AI-generated voices as "artificial" under existing telemarketing consent rules, which matters if you're using cloned voices in any outbound calling context. In the EU, voice cloning falls under both GDPR and the EU's AI regulation, which requires consent for creating, storing, and distributing a cloned voice.
Practically, this means: don't clone a voice without documented consent from the person it belongs to, check each provider's specific terms on commercial use of cloned voices (they vary more than you'd expect), and if you're operating across multiple countries, assume the strictest applicable jurisdiction's rules apply rather than picking whichever is most convenient.
Where Voice Fits Alongside AI Video and Music
Voice rarely stands alone in a real production pipeline. Narration and dubbing pair naturally with AI-generated video, either as a voiceover layered onto a generated clip or as a translated dub for a piece originally recorded in another language. Background music pairs with both, filling out a short-form video or ad with a soundtrack instead of leaving it silent. A unified API that covers image, video, and music generation alongside voice makes this kind of pipeline easier to build, since you're not stitching together separate vendors, API keys, and billing dashboards for what's ultimately one output. If you're already generating video through something like Apiframe's AI video API, adding narration or dubbing to the same pipeline is a smaller lift than bringing in a completely separate voice vendor.
You can also see how Apiframe approaches other media workflows in its text-to-video Python tutorial, which walks through authentication, sending a generation request, polling for completion, and handling the resulting video.
For music, Apiframe's Best AI Music Generation APIs in 2026 compares services such as Suno, Udio, ElevenLabs Music, and other options based on quality, pricing, and licensing.
FAQ
What is the cheapest text-to-speech API?
Amazon Polly is among the cheapest at around $4 per million characters for standard voices, with Deepgram Aura-2 and Cartesia's lower tiers also landing well under ElevenLabs' rates. Cheapest isn't always best, though; check naturalness and voice cloning support against what you actually need.
Which text-to-speech API sounds most natural?
ElevenLabs consistently ranks highest for naturalness and expressiveness in independent comparisons, particularly for cloned or emotionally expressive voices. For straightforward narration, several competitors are close enough that the difference is hard to hear outside a direct side-by-side test.
Do these APIs support real-time streaming?
Most of the major providers do to some degree, but Cartesia is specifically built around it, with the lowest published latency in the category. If real-time responsiveness is a hard requirement, test latency yourself rather than trusting a headline number, since real-world performance depends on your own network path and text length.
Is there a free text-to-speech API tier?
Most providers offer a limited free tier or trial credit for testing, ElevenLabs, PlayHT, and Cartesia all do, but none are meant for production use beyond light testing volumes.
