feat: Synthetic noise generation .apr model (WASM-first)

[noahgift]

Summary

Build a .apr model for generating synthetic audio noise (white, brown, pink, etc.) using ML/DL techniques. WASM-first approach enables browser-based interactive demos and CLI execution.

Motivation

• Demonstrate .apr format capabilities with audio generation
• WASM-first architecture for instant web demos
• Showcase ML-based audio synthesis vs traditional DSP
• Provide infinitely customizable noise generation beyond static presets

Requirements

Core Features

• .apr model that generates continuous noise streams
• WASM build for browser deployment
• CLI execution via aprender run noise.apr
• Real-time parameter adjustment (no model reload)

Noise Types (via ML generation)

Type	Characteristics
White	Flat frequency spectrum
Brown/Red	-6 dB/octave, deeper
Pink	-3 dB/octave, balanced
Blue	+3 dB/octave, brighter
Violet	+6 dB/octave
Custom	User-defined spectral profile

Customization Parameters

• Spectral shape: Arbitrary frequency curve (not just fixed slopes)
• Modulation: LFO on amplitude, frequency bands
• Binaural: Optional stereo offset for binaural beats
• Texture: Grain size, smoothness, rhythmic patterns
• Blend: Mix multiple noise profiles

WASM Demo Interface

┌─────────────────────────────────────────┐
│  🎵 Noise Generator                     │
├─────────────────────────────────────────┤
│  Type: [White ▾] [Brown] [Pink] [Custom]│
│                                         │
│  Spectral Curve:  ═══════════╲          │
│                              ╲          │
│  Depth:     ████████░░ 80%              │
│  Texture:   ██████░░░░ 60%              │
│  Modulation:███░░░░░░░ 30%              │
│                                         │
│  [▶ Play]  [⏸ Pause]  [💾 Export .apr]  │
└─────────────────────────────────────────┘

Technical Approach

Option A: Diffusion-based

• Small diffusion model for audio generation
• Spectral conditioning for noise color
• Higher quality, more compute

Option B: Autoencoder + Decoder

• VAE-style latent space for noise characteristics
• Fast inference, lower memory
• Easier WASM deployment

Option C: Neural vocoder adaptation

• WaveNet/WaveGrad style generation
• Real-time streaming possible
• Most complex

Recommended: Start with Option B for WASM feasibility, iterate to A/C.

Dependencies

• aprender::audio (mel, resample - already exists)
• trueno (SIMD acceleration)
• wasm-bindgen / web-sys (browser integration)
• Web Audio API for playback

Acceptance Criteria

1. cargo build --target wasm32-unknown-unknown --features audio succeeds
2. Demo page loads model and generates audio in <2s
3. Parameter changes reflect in audio within 50ms
4. Model size <5MB for fast web loading
5. Works offline after initial load

References

• WaveNet - Autoregressive audio
• Diffusion models for audio
• Existing aprender::audio::mel for spectral processing