Products · Audio · Showcase
One engine, many voices.
Thirty-six directed voices, one engine. Every card below plays its take.
Enter the showcase
A grand host invites you in. Press play — then take the voices one by one, starting with a feeling.
Read the invitation script
Control room
One take, every trick — a showman host demonstrates each capability while narrating what it does. Generated once with the frontier model.
The guided tour — one take, every trick
Seventy-one seconds, nine tricks, two voices. The voice demonstrates each capability while narrating what it does. Press play.
- plain direction — briefed like an actor, performed vivid.
- <emphasis> — surgical weight on a single word.
- <whisper> — dynamics, close then full voice.
- [small laugh] — written laughter, never recorded.
- <fast> / <slow> — the pace instrument.
- phonemes — pronunciation pinned exactly.
- <french> — language switch, experimental.
- speaker tags — a directed duo in one take.
Same voice, four emotions
One voice, four ways — and four different texts, so the contrast is audible, not subtle. Each card plays its directed take.
Picked for the use case
Clear voices for clarity-critical slots: a crisp measured documentary read, and an Arabic voice reading English.
Newsroom
Same MSA lede on the official voice and a usage-ranked community newsreader — both fully diacritized, exact readings preserved — plus an English markets read.
Advertising
A warm French sell and an epic English sell, one line each.
Audiobook, with expression and effects
Narration where the feelings and the sound effects ride inside the text: pauses, whispers, sighs, and a two-sentence sad-to-happy arc. Parsing verified — every tag below reaches the engine verbatim (see the evidence in the research notes).
Game characters
Community archetypes with unordinary timbre — a raspy wizard, a deadpan robot — beside two voices designed from a single sentence: a gravelly chieftain and a bright pixie.
Podcast
Two voices, one job: fiction teasers and studio skits with speaker tags.
Customer support
A helpline that never sleeps: agent and caller, directed in a single take.
Video automation
Three platform pacings: a sixty-second YouTube promise, a brisk TikTok hook from a clear announcer with an emphasis tag, and a calm Sunday reel.
Entertainment
A wedding-toast tale that laughs mid-sentence — vocal effects, not post-production.
Exotic shelf
Top voices per language — then the odd heads: a high-meets-deep Monster, a raspy Goblin, a Gothic literary creature, and a breathy ghost. Unordinary timbre and pitch, same API call shape.
Languages
The same welcome, performed natively in Spanish, German, and Italian.
Clone: source and result
The training sample is itself synthetic narration (no human voice involved), cloned with explicit consent. Compare the source with the clone reading the anchor sentence.
Designed voice
Built from a single instruction sentence — no sample audio at all — then saved as a persistent voice.
Bonus: the same engine listens
The 20-second clone source, transcribed back by the ASR lane with word timestamps and language detection — and every transcript below plays its actual audio.
Go deeper
This page is a peek preview. The full stories live in the articles.
Devlog: voice, live on a laptop
How the realtime voice agent runs locally — the engine behind these takes.
Why every word needs a timestamp
The timing craft that keeps every transcript honest, from studio to ASR.
What pre-launch means here
Waitlist, roadmap, and what is verifiable today versus promised.
Get access
Hosted access is onboarding from the waitlist.