Introduction

Inworld just shipped Reasoning for Realtime TTS-2, an opt-in switch that lets its closed, API-only text-to-speech model decide how a line should sound before it says it. You send the text plus a bit of scene context, TTS-2 writes itself a short acting note like "warm, delighted, speaking through a smile," then performs it. The note comes back in the API response, so you can see exactly what direction it chose.

It runs behind the same inworld-tts-2 model ID at the same price, and you can try it today in the Inworld TTS playground (where it's on by default as a beta) or through the API. Inworld

on October 29, calling it the first TTS model that can steer itself from context. I can't verify "first," but the idea is genuinely useful.

What Shipped

From Inworld's launch post and steering docs:

  • Reasoning: a small language model inside the same TTS-2 request reads your line and context, writes a free-form delivery direction, then TTS-2 speaks it. No extra service to call or host.
  • Context field: reasoningContext holds up to 600 characters describing the scene or character. It's never spoken and isn't billed.
  • Conversation memory: earlier lines go in previousRequests so the read fits the flow of the conversation.
  • Visible output: the chosen direction comes back as reasoningInstruction (on the first message only when streaming).
  • You stay in charge: your own inline [tags] always win. On async jobs, your first tag ends Reasoning for the rest of that job.
  • Never blocks: if Reasoning can't write a direction in time, the line is spoken without one. It does add a short step before first audio on regular and streaming requests.
  • Where it works now: standard synthesis, streaming, and async jobs on inworld-tts-2. WebSocket streaming, duplex (including the Realtime API) and batch are "coming soon." It doesn't run on TTS-2 Flash, which doesn't support steering.

The Model Update Riding Along

The same release also swaps in an updated TTS-2 under the same ID, rolling out by region with no migration:

  • Speed: about 13% faster generation on short lines and about 9% on mid-length ones, per Inworld's comparison with the previous version.
  • Voice matching: closer clones in Stable mode, especially for voices that were hard to copy, and less drift over long duplex sessions.
  • Fixes: fewer skipped words in long passages and fewer dropped end consonants like /p/, /t/ and /k/.
  • Creative mode: more natural delivery with light built-in steering.

Inworld also cites Cekura's independent benchmark: Realtime TTS-2 Flash had a 46 ms median to first audible sound (fastest of 27 models in its non-streaming test), and TTS-2 read 93.4% of tricky items correctly, like codes, addresses and medical terms. Those Cekura numbers were measured before this update and without Reasoning.

Pricing

Nothing changes. From Inworld's pricing page:

  • On-Demand: Realtime TTS-2 at $25 per 1M characters, with up to 70 minutes of TTS included free.
  • Creator ($25/mo): $20 per 1M characters.
  • Builder ($100/mo): $17.50 per 1M characters.
  • Enterprise: as low as $5 per 1M characters.

How to Try It Today

The quickest way is the TTS playground on Inworld's TTS page, where Speech Reasoning is switched on by default. For the API, add one field to a normal TTS-2 request:

curl -X POST https://api.inworld.ai/tts/v1/voice \
  -H "Authorization: Basic $INWORLD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Don'\''t open that door.",
    "voiceId": "Dennis",
    "modelId": "inworld-tts-2",
    "synthesisContext": {
      "reasoning": "REASONING_ON",
      "reasoningContext": "Protagonist in a psychological thriller, alone in the house at night.",
      "previousRequests": [{ "text": "Did you hear that?" }]
    }
  }'

The response includes your audio plus a reasoningInstruction string with the direction TTS-2 picked.

Pro Tip

Log reasoningInstruction next to every line you generate. When a read comes out wrong, you'll usually find the direction was reasonable for the context you gave it, so fix the context instead of hand-tagging the line. Keep context about the scene and character, and put earlier dialogue in previousRequests rather than stuffing it into the 600-character field.

Why It Matters for Audio Creators

If you've ever built a talking character, you know the annoying part isn't the voice. It's getting your LLM to pick the right emotion tag for every reply, and maintaining the prompt rules that make it do that. Inworld's pitch is that a fixed tag like [happy] sounds the same in every scene, while a written direction can set tone, emotion and pacing together.

That's a real fit for audiobooks, generated stories and lessons, companion voice messages, and pre-generated game dialogue, which are exactly the cases Inworld says Reasoning supports at launch. Live voice agents on the Realtime API will have to wait for duplex support.

Conclusion

Inworld's Realtime TTS-2 can now write its own acting note from your context before it speaks, and it tells you what note it chose. It's closed and API-only, opt-in with one field, and costs the same $25 per 1M characters on demand. Try it in the playground with one line and two different contexts, and listen for the difference.

—Anabel ♡