@threejson/audio-kit

Optional local audio primitives for ThreeJSON or other applications. Importing this package does not download speech models, create an AudioContext or call a provider.

import { synthesizeScore, encodeWav } from '@threejson/audio-kit';
const pcm = synthesizeScore({ version: 1, tracks: [{ id: 'melody', instrument: 'bell',
  notes: [{ id: 'note', tick: 0, duration: 960, pitch: 60 }]
}] });
const wav = encodeWav(pcm);

Use createScoreRenderer and createPcmMixer for bounded-block synthesis. createPcmPlayback schedules the same blocks on a WebAudio clock. Basic electronic instruments are not natural singing.

Optional subpaths: /models (explicit download/import, SHA-256 verification, OPFS/IndexedDB, leases), /node-models (user cache), /worker-producer, /sherpa, /soundfont.

/models exports getBuiltinAudioModels and createLocalSpeechProducer({modelManager}): a pinned MeloTTS Chinese + Sherpa-ONNX single-thread WASM preview, about 71 MiB, loaded only after explicit installation. Actual Chinese PCM synthesis was verified in local Edge; this is not a claim of professional voice quality or mobile performance. Kokoro and other voices still need adapters. Weights are not shipped or downloaded by npm installation.

analyzePcm returns RMS, peak, clipping and a sampled envelope. The deterministic mixer supports gain/pan automation and music ducking keyed to narration or explicit clip IDs. Generated WAVs are portable; listeners do not need the original voice model.

中文指南 · English guide