Inworld Realtime TTS-2 is a hosted text-to-speech API with a no-code playground: pick a stock voice, type text, get audio back in seconds, with instant voice cloning and 200+ languages. Pick it for a quick spoken clip or a streaming voice feature. It only speaks text; it doesn't listen, and per-character billing needs a plan check.
Other good fits
Adding streaming, low-latency speech output to an app that already handles the conversation logic.
+ Pros
Read pros
Inworld Realtime TTS-2 (model ID inworld-tts-2) turns text into speech through a browser playground or an API. It supports natural-language steering ("say this more cheerfully"), 200+ languages, and instant voice cloning from a short sample. A faster, cheaper sibling, TTS-2 Flash, trades some quality for roughly 5x lower latency, so you can pick quality or speed for the same text.
Pick it when…
You want to try a stock voice on a short piece of text without writing any code.
You're adding a streaming voice output to an app and want a documented, low-latency endpoint.
You need voice cloning from a short sample rather than only pre-built voices.
− Cons
Read cons
TTS-2 only produces speech from text: it has no speech recognition and no conversation logic, so it's not a drop-in voice assistant. The synchronous endpoint accepts at most 2,000 characters per request; longer text needs the streaming, async or batch options instead. The docs describe generation history as stored per device and workspace in the playground, but don't spell out a full retention or training policy for API input and generated audio, so check Inworld's data-processing terms before sending anything sensitive.
Pick something else when…
You need speech recognition too.Deepgram covers the listening half; pair it with a TTS API for the reply.
You want a different realtime speech API to compare against.Cartesia is a similar hosted option worth trying side by side.
Everything must run on your own machine. Use a local runtime such as Ollama for local language-model work; it's not a hosted TTS replacement, but it keeps data off the network entirely for the language side of a voice project.
Avoid if
Your material must stay on your own machine; this is a cloud API with per-character billing.
Connection
The playground lives in the Inworld Portal and needs no code. API calls use an Inworld API key; direct server-to-server requests use Basic authentication, while browser and mobile apps should go through the documented backend token flow rather than embedding a key in client code. The published latency figure (100 ms time to first audio byte for TTS-2, 20 ms for Flash) is measured server-side and excludes your own network time.
There's nothing to install; the running effort is watching character counts, since a long document adds up fast at these per-character rates, and rotating the API key if it's ever exposed.
Data handling
The Inworld Portal stores generation history locally to each device and workspace. The docs checked here don't lay out a complete data-retention, training-use or regional-processing policy for API text input or generated audio, so treat that as unconfirmed and check Inworld's current terms before sending sensitive or personal content.
First request
Documentation-based suggestion — not run here.
Make a spoken invitation for a lab open house.
Open the TTS Playground and pick a stock voice that fits a friendly, informal announcement.
Type a two-sentence invitation: what's happening, when, and where.
Generate the clip, listen to it, and try one steering instruction, for example "more upbeat".
Download the audio file.
Done when you have a downloadable clip that sounds natural and says exactly what you typed, with no cloned voice or personal data involved.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Rests on Inworld's TTS model documentation, playground and quickstart docs, and pricing page, all checked 23 September 2026. No account, API key, generation, cloning or billing was exercised for this review.
Inworld TTS models · Sep 23, 2026 — Model ID inworld-tts-2, natural-language steering, 200+ languages, voice cloning, TTS-2 vs Flash latency figures.
Inworld TTS Playground · Sep 23, 2026 — No-code voice selection, generation and download; per-device/workspace generation history.
Inworld developer quickstart · Sep 23, 2026 — API key handling and browser/mobile backend token flow.
Inworld synthesize speech · Sep 23, 2026 — 2,000-character limit on the synchronous endpoint; streaming, async and batch alternatives.
Inworld pricing · Sep 23, 2026 — Plan table with monthly cost, included credits and per-million-character TTS-2 rate.
Further review notes & alternatives
Alternatives and trade-offs
Cartesia wins when you want another realtime speech API to compare on latency and voice quality.
Deepgram wins when you need speech-to-text as well, since TTS-2 doesn't listen.
ElevenLabs wins when voice-cloning polish and a larger voice library matter more than Inworld's per-character price.
