Why architecture matters here
Music generation fails on rights (models trained on copyrighted music), quality (uncanny synth), and control (users can't steer). The architecture matters because rights + watermark + consent must be built in.
The architecture: every piece explained
The top strip is the pipeline. Text/MIDI prompt conditions. Audio tokenizer — Encodec, SoundStream — encodes to discrete tokens. Acoustic transformer predicts tokens. Vocoder / detokenizer reconstructs audio.
The middle row is control + safeguards. Style conditioning selects genre + mood + instrument. Multitrack support generates stems separately. Watermark for AI detection. Rights + consent — training data provenance + artist voice consent.
The lower rows are ops. Streaming generates as it plays. Metrics — MOS + genre accuracy. Ops handles versioning, eval, licensing, safeguards.
End-to-end flow
End-to-end: user prompt "upbeat jazz piano solo". Tokenizer encodes silence starter. Acoustic model conditioned on prompt predicts tokens. Vocoder streams audio. Watermark embedded. Rights check confirms training corpus was licensed. User downloads with license.