Directing voices instead of tagging them.

Task progress, measured: a voice showcase built take by take — generating, measuring, and keeping only what survived. If you stop here: every control below was A/B-tested on identical sentences, and the audios speak for themselves in our voice showcase.

It started with square brackets. An emotion tag in front of a sentence — happy, sad, a whisper — and the voice obeys. Or claims to: three takes came back sounding alike, others came back visibly wrong. So takes were regenerated over and over, text and tags refined each round, until speech matched script — with durations and transcripts as the ledger. The decisive tests used identical sentences: same voice, only the tag varies.

Every tag on trial

The results read like a spec sheet. An emphasis marker stretches delivery by 23%. A soft-tone marker shortens it by 17%. A laugh tag adds more than a third in length — it genuinely laughs. A sadness tag barely moves timing at all; it colors instead of pacing. A cough adds almost a full second and comes back in transcription as a cleared throat.

Two properties held everywhere we looked. First, tags shape delivery but never leak into words — transcripts come back clean. Second, finer tools do finer work: a sentence-level emphasis is a blunt instrument next to a tag that stresses one single word.

Tags shape delivery. They never leak into words.

When effects eat the script

Not every direction behaves. A pause marker the engine never documented came back as the sound of something breaking. One take ran 47 seconds with 44 seconds of dead silence. A crowd-cheer opening swallowed its own script: 38 seconds of audio, 21 of 65 words spoken, the rest lost to ambience.

The rule we took from the wreckage: sound effects as short accents only, never as atmosphere. A laugh lands. A crowd bed drowns. We regenerated the broken takes, verified each by transcription, and kept only the ones whose words all arrived.

Two engines

We now think in two tiers. The stable default speaks tagged takes — it carries the whole demo. The frontier preview takes plain-language direction, scopes single words, even switches languages mid-line. Run side by side on identical scripts, the frontier consistently acts bigger: longer takes, wider pacing.

The frontier is where our most directed pieces come from. The stable default is what visitors play with. Knowing which job needs which engine is most of the craft. Where this all ships is covered in what pre-launch means here.

Whose voice

A word on casting. Every voice on the showcase is a generic archetype — storytellers, a trailer voice, a robot, a wizard — Fictional personas are designed from a sentence instead. Voice cloning itself stays behind an explicit consent gate.

Hear it

Forty-two takes, fourteen sections: emotions, newsrooms, audiobooks, game characters, podcasts, automation, languages, and more. Start with the guided tour — seventy-one seconds, nine tricks, two voices — then take the rest one by one.