Under the Hood: Deep Cuts Architecture

Deep Cuts is built with a highly optimized, fully local native desktop stack. By eliminating external cloud processes and background services, it executes digital signal processing (DSP) algorithms and neural network models entirely in-process on consumer CPUs and Apple Silicon hardware.

Core Tech Stack

• Frontend: Svelte 5 (in static SPA mode), SvelteKit, and Vite. Utilizes D3.js for canvas visualizations and WaveSurfer.js for audio waveforms and spectrogram elements.
• Backend Core: Tauri v2 and native Rust.
• Local Database & Vector Index: SQLite (rusqlite) compiled with sqlite-vec virtual table configurations to enable fast, local K-Nearest Neighbors (KNN) vector calculations.
• Local Neural Inference: ort (native ONNX Runtime bindings) runs CLAP and Essentia classifiers.
• Local Multimodal LLM: Bundled llama-server (llama.cpp) binary resolves local GGUF models (Qwen2-Audio) for offline, sandboxed audio Q&A sessions.

---

Audio Analysis Pipeline

Deep Cuts processes watched folders recursively using a concurrent, spool-based analysis engine operating in dependency order across worker threads (num_cpus / 2):

1. BPM & Temporal Analysis: Spectral-flux onset envelope tracking with autocorrelation and parabolic sub-sample refinement (40–210 BPM range).
2. Double-Pass Correction:
   • Genre metadata coarse pass to resolve double-time or half-time errors.
   • Secondary sweep matching against Essentia classifier outputs for alignment.
3. Key & Scale Detection: Chromagram constructed via FFT, HPCP harmonic suppression, and Krumhansl-Schmuckler profile correlation.
4. Loudness (EBU R128): Analyzes integrated loudness (LUFS) and loudness range (LRA) using the native ebur128 library.
5. Essentia Genre & Mood Classifiers: Discogs-Effnet ONNX models evaluate 400 genre classes, vocal vs. instrumental probabilities, and seven mood axes (happy, sad, aggressive, relaxed, party, acoustic, electronic).
6. LAION CLAP Embeddings: Extracts 512-dimensional audio embeddings from high-energy (loudest) 10-second audio windows to ensure silent intros do not pollute vector indices.
7. Description Embeddings: Encodes local description tags using the all-MiniLM-L6-v2 ONNX model into 384-dimensional vectors for natural language semantic search.

---

Non-Destructive Caching (.dc.json Sidecars)

To protect pristine audio files from metadata modification, Deep Cuts writes computed analysis results into lightweight .dc.json sidecar files stored next to each track (optional, toggled via Settings).

• Relocation Safety: If files are moved or directories renamed, re-indexing is near-instantaneous as the application reads existing sidecars instead of re-running the CPU-intensive analysis pipeline.