Inworld Realtime TTS-2

Introduction

Most text-to-speech models read words without considering how those words should sound. A farewell and a punchline come out with the same pitch, the same pacing, the same flat affect. The model knows what to say but has no concept of how to say it.

TTS-2 fixes this with voice direction: natural-language instructions, written in square brackets, that tell the model how to deliver a line. You write them the way you'd write a stage direction for an actor. There are no preset emotion slots, no dropdown menus, no sliders. If you can describe the delivery in plain English, the model can perform it.

How steering tags work

A steering tag is a natural-language instruction enclosed in square brackets, placed before the text it applies to. The model reads the tag, adjusts its delivery, and speaks the text that follows.

The syntax is straightforward:

[instruction] Text to speak.

The tag applies forward until the model reaches the next tag or the end of the input. You don't need to close it or reset it. If the mood shifts mid-text, drop a new tag at the transition point:

[speak cheerfully] Great news, the build passed! [lower voice, more serious] But we need to talk about the memory leak.

Here's the full API request for that example:

[
  {
    "taskType": "audioInference",
    "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "model": "inworld:tts@2",
    "speech": {
      "text": "[speak cheerfully] Great news, the build passed! [lower voice, more serious] But we need to talk about the memory leak.",
      "voice": "Sarah"
    }
  }
]

{
  "data": [
    {
      "taskType": "audioInference",
      "audioUUID": "f1e2d3c4-b5a6-7890-1234-567890abcdef",
      "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
      "audioURL": "https://aud.runware.ai/audio/f1e2d3c4-b5a6-7890-1234-567890abcdef.mp3"
    }
  ]
}

Writing effective directions

Without any steering tags, the model reads text in a neutral, default tone. Adding a direction tag transforms the delivery. Compare the same line read flat versus with a detailed direction:

The directed version layers mood, rhythm, pitch, and vocal mode into a single instruction. Think of it as the difference between handing an actor a script versus giving them a scene brief. The more dimensions you describe, the more specific the performance.

Direction categories

Steering tags aren't limited to emotions. You can control multiple dimensions of delivery:

[say excitedly] We just hit a million users!
[sound concerned, speaking carefully] I'm not sure that's the right approach.
[sound terrified, barely holding it together] Did you hear that?

[articulate clearly, with deliberate pauses] The API key goes in the Authorization header.
[very fast, urgent] We need to push the hotfix now, production is down.
[very slow, measured] Let me explain this one more time.

[very quiet, almost a whisper] Don't tell anyone, but the feature ships tomorrow.
[say in a low tone, gravelly] This is your last warning.
[say in a high pitch, bright and cheerful] Good morning! Ready to get started?

[whisper in a hushed style] I think someone's listening.
[give a nasal quality, slightly annoyed] Ugh, another meeting.

Inline non-verbals

You can also insert non-verbal sounds at specific points in the text. These are placed inline, exactly where the sound should occur:

Wait, you actually did that? [laugh] That's wild.
[sigh] I don't know. It's been one of those weeks.
Okay, let me think about this for a second. [breathe] Right. Here's what we do.

The available non-verbals are [laugh], [sigh], [breathe], [clear throat], [cough], and [yawn]. They work as audio events, not pronounced words. The model inserts the sound at that position and continues speaking.

You can combine non-verbals with steering tags. The tag sets the delivery, the non-verbal inserts a specific moment:

Emphasis and pacing

Two text-level techniques complement steering tags and work in any TTS-2 request, with or without bracket tags.

I told you NOT to do that.
Your order will arrive by FRIDAY.
That is AbsoLUTEly correct.

I thought it would work, but... I'm not so sure anymore.
Stop. Think. Then act.
Well, when you put it that way, I suppose you might have a point.

Common mistakes

Tips for best results

1. Write stage directions, not keywords. "Say this like you're apologizing to someone you care about" outperforms "apologetic". The model responds to prose, so give it a scene.

2. Combine at least two dimensions. Mood alone is generic. Mood plus pacing, or mood plus volume, produces a more specific delivery. "Sad" is vague. "Sad, quiet, with long pauses" is a read.

3. Place non-verbals where a human would do them. A [sigh] at the start of a resigned sentence, a [laugh] after a punchline. Don't scatter them randomly.

4. Use capitalization for one or two words per sentence at most. Emphasis works because it's selective. Capitalizing half the sentence removes the contrast that makes it effective.

5. Test with the same voice. Different voices respond to the same direction with different intensity. A direction that sounds perfect on Sarah might feel understated on Ethan. Pick your voice first, then tune the directions to it.

6. Let the model breathe. Longer input with natural punctuation sounds better than short fragments stitched together. Give the model a full thought to work with, not a word at a time.