CassetteAI describes itself as real-time audio machine learning running on the edge, and the numbers it publishes are the core of the pitch: roughly 23 ms first-sample latency, a 30-second music sample rendered in under two seconds, a full three-minute track in under ten, and sound effects up to 30 seconds returned in about one second. Output is 44.1 kHz stereo — the reference rate CDs and DAWs expect — and the models expose deterministic seeds so a generation can be reproduced.
The single SDK covers three modalities. The music model generates one- to four-minute tracks with drums, bass, harmony and lead from a text prompt describing mood, genre or a reference, streamed into an app while the user plays. The sound-effects model produces loop-safe clips from descriptions like "heavy door slamming in a cathedral" or "retro power-up", short enough and quick enough to re-roll per frame in a game or AR/VR experience. Text-to-speech is listed as coming soon, built around streaming output, sub-second first phoneme and zero-shot voice cloning from a ten-second sample.
Commercially, CassetteAI is metered rather than tiered: it publishes $0.02 per output minute for music and a flat $0.01 per generation for sound effects, with no monthly commit and no developer seats. There is a no-signup live playground on the site for trying prompts, and the company invites teams needing dedicated capacity, volume pricing or on-device licensing to contact it directly.
