# 🚀 Deep Cuts v0.1.0 — First Binary Release

Welcome to the first official binary release of **Deep Cuts**! This release marks a major milestone: a 100% offline, private, and zero-dependency studio audio analysis and music intelligence desktop application designed to run locally and sandboxed on macOS.

[Download it from GitHub](https://github.com/robertolupi/deep-cuts/releases/tag/v0.1.0)

%[https://youtu.be/bev7k1K8cmY] 

* * *

### ✨ Core Features

#### 🎛️ 100% Offline Digital Signal Processing (DSP)

*   **Precision BPM Detection**: Spectral-flux onset envelope tracking with autocorrelation and parabolic sub-sample refinement (40–210 BPM range).
    
*   **BPM Double/Half-Time Correction**: Automatic dual-pass sweeps using coarse metadata genres and high-fidelity machine learning style classifiers to resolve tempo scale errors.
    
*   **Key & Scale Detection**: Harmonic Pitch Class Profiles (HPCP) suppress irrelevant overtones, correlating with Krumhansl-Schmuckler profiles for accurate key detection.
    
*   **EBU R128 Loudness**: Calculates precise integrated loudness (LUFS) and Loudness Range (LRA) to match studio standards.
    
*   **Waveform Profiles**: Extracts 128-point RMS energy profiles for rapid visual rendering.
    

#### 🧠 Local AI & Machine Learning Indexing

*   **Essentia Classifier (Discogs-Effnet)**: Offline neural network style analysis predicting **400 hierarchical genre classes**, vocal/instrumental likelihood, and **7 mood axes** (happy, sad, aggressive, relaxed, party, acoustic, electronic).
    
*   **CLAP Acoustic Embeddings**: LAION CLAP ONNX model extracts deep semantic embeddings from 10-second windows, enabling instantaneous **acoustic similarity (KNN) search** via a local `sqlite-vec` virtual table.
    
*   **Description Embeddings**: Pinned `all-MiniLM-L6-v2` ONNX model encodes natural language text and generated descriptions to facilitate semantic search.
    

#### 💬 Local Multimodal AI Chat

*   **Conversational Track Q&A**: A dedicated Chat tab in the Track Detail Pane allowing producers to converse directly with their audio files using **Qwen2-Audio-7B-Instruct** running completely locally.
    
*   **Sonic Arrangements & Troubleshooting**: Ask questions like: *"Why does this mix sound muddy?"*, *"What is the song structure?"*, or *"Give me EQ tips for this master."*
    
*   **Context-Smart Payload**: Waveform metadata is attached only on the first turn; subsequent turns default to lightning-fast, text-only conversation to preserve context length.
    
*   **WaveSurfer Region Selector**: Interactive waveform drag-and-drop region selection to target analysis on specific song sections (intro, chorus, drop) without blowing context windows.
    
*   **Real-time SSE Token Streaming**: Full Server-Sent Events integration for instantaneous token streaming and highly responsive, snappy chat feedback.
    

#### 🗺️ Interactive Music Map (UMAP)

*   **2D Audio Projection**: High-performance Rust-native UMAP dimensionality reduction projects your entire CLAP-embedded collection onto an interactive 2D canvas.
    
*   **Filter-Aware Canvas**: Smooth D3-driven zooming, panning, and legend rendering instantly reflect active sidebar filters (BPM, Key, Genre, Monitored folders, Vocals).
    
*   **Neighborhood KNN Inspector**: Click any coordinate to reveal and scrub through the nearest 10 acoustically similar reference tracks in your library.
    

#### 📁 Zero-Config Staging & Bundling

*   **Fully Self-Contained Sidecar**: The app bundles the local LLM runtime (`llama-server`) alongside all its modular dynamic dependencies (`libggml`, `libllama`, `libllama-server-impl`, etc.) out-of-the-box, ensuring relocatability and preventing system linker mismatches.
    
*   **In-App Model Downloader & Verification**: Dynamic folder configuration allows users to store models on custom external drives or local sandboxed application data directories. Features dynamic verify-and-resume downloads for heavy model checkpoints directly inside the settings drawer.
    
*   **Sidecar JSON Persistence (**`.dc.json`**)**: Auto-writes lightweight sidecars next to your files. If you move your library or reinstall the app, re-indexing takes fractions of a second by directly restoring computed features from sidecar files.
    

* * *

### 🔧 Installation

1.  Download [`Deep Cuts_0.1.0_aarch64.dmg`](https://github.com/robertolupi/deep-cuts/releases/tag/v0.1.0) for Apple Silicon Macs.
    
2.  Drag and drop **Deep Cuts.app** into your `Applications` folder.
    
3.  Open Settings (gear icon), select your **Model Folder** (or use the default sandboxed folder), and click **Manage Models** to fetch your local weights with a single click!