Realtime TTS API

A TTS (text-to-speech) API converts written text into spoken audio via HTTP requests. You send text to an endpoint and receive audio back-as MP3, WAV, or streaming chunks. TTS APIs power voice assistants, audiobook generation, accessibility features, video narration, and AI voice agents. Modern APIs like Inworld offer realtime streaming, instant voice cloning, and support for 200+ languages with Realtime TTS-2.

Realtime TTS is top-ranked on independent speech benchmarks, including the Artificial Analysis TTS Arena, a public leaderboard based on blind listening tests by thousands of real users. It combines top-ranked voice quality with sub-100ms latency and pricing that scales down to as low as $5 per million characters. For most applications, start with Realtime TTS-2, the flagship for expressiveness and steering. Pick Realtime TTS-2 Flash for the latency and cost floor.

Realtime TTS is top-ranked on Artificial Analysis. Realtime TTS-2 Flash delivers 20 ms TTFB down to $5/1M characters, while Realtime TTS-2 pairs 100 ms TTFB with the best quality and natural-language steering. Inworld also provides WebSocket streaming for realtime voice agents, on-premise deployment, and instant voice cloning from 5 to 15 seconds of audio.

Realtime TTS-2 runs as low as $5 per million characters at scale, about $0.005 per minute of audio. Realtime TTS-2 Flash starts at $15 per million and falls to $7. Rates fall as your total spend grows, and spend on TTS counts toward one combined commit with the Router, speech-to-text, and compute. Exact tiers and a cost calculator are at inworld.ai/pricing.

Make a POST request to the TTS endpoint with your text, voice ID, and model. You'll receive audio back as base64-encoded data. For realtime apps, use WebSocket streaming to receive audio chunks as text is generated. Works with Python, Node.js, curl, or any HTTP client. Generate an API key and check the Developer Quickstart for code examples.

Inworld Realtime TTS-2 Flash delivers 20 ms TTFB*, and Realtime TTS-2 runs 100 ms while adding natural-language steering. Both support native WebSocket streaming where audio generates instantly with no buffering delay. This makes Inworld ideal for voice agents, interactive media, and any app where response time is critical.

*P90 time to first audio byte, measured server-side — excludes network latency.

Realtime TTS is top-ranked for voice quality on Artificial Analysis, based on blind tests by thousands of real users. Realtime TTS-2 is optimized to minimize hallucinations, word cutoffs, and audio artifacts, and holds one voice identity across every language it speaks. Try it in the Playground.

Yes. Inworld offers instant voice cloning: upload 5 to 15 seconds of audio and get a unique voice ID to use in TTS requests immediately. For maximum fidelity, professional voice cloning uses 30+ minutes of clean audio (minimum 5 minutes, 20+ minutes recommended). Contact sales for professional voice cloning.

Realtime TTS-2 and Realtime TTS-2 Flash support 200+ languages with one voice identity preserved across every language and mid-utterance switching inside a single generation. Use the List Voices endpoint to filter available voices by language.

REST returns complete audio after all text is processed-best for batch jobs and short text. HTTP streaming returns audio chunks progressively for faster playback start. WebSocket maintains a persistent bidirectional connection, ideal for voice agents where text is generated incrementally by an LLM. Inworld supports all three from a single API.

Our published latency figures are on-server inference time, which measures how long our models take to generate the first audio chunk once the request reaches our servers:

Your end-to-end latency also includes network round-trip time between your application and our servers.