voice steering

Voice steering lets you tell an AI voice how to "act" in plain language, and it follows the direction. It's similar to giving a voice actor stage directions like "whisper this part" or "say this line angrily" — except instead of a human actor, an AI picks up the instruction on the spot and applies it to the voice.

Traditional text-to-speech usually reads everything in one fixed tone. If you wanted a sad voice versus a cheerful one, you had to pick a different preset from the start. With voice steering, the same sentence can sound completely different just by changing the prompt to something like "say it in a tired, low voice." Much like large language models follow natural-language instructions, speech models with voice steering let you control tone, emotion, and delivery through instructions as well.

Switching languages or dialects while keeping the same speaker identity is often introduced together with this feature. It's used in things like call-center bots that shift tone depending on the situation, or audiobook narration that expresses different emotions scene by scene.

Inworld AI introduced its text-to-speech model Realtime TTS-2, saying that "with voice steering, you can adjust the tone or emotion of a voice on the fly using natural language, much like giving acting direction." When this term comes up, it doesn't mean changing the voice itself (the speaker's identity, such as gender or age) — it specifically means adjusting the same speaker's tone, emotion, and delivery style through instructions.

Try entering the same sentence twice into a text-to-speech tool and compare the results.

1. Plain text only: Today's meeting has been canceled.
2. With an instruction added: (in a tired, low voice, sounding apologetic) Today's meeting has been canceled.
3. With a different instruction: (bright and upbeat, like an announcement) Today's meeting has been canceled.

Listening to how much the tone and emotion change with the exact same text gives you an intuitive sense of what voice steering actually does.