Abstract

AudioCrucible is a desktop application that uses on-device AI to automatically classify and rename sound effect files according to the Universal Category System (UCS) standard. It combines a neural audio encoder with a local language model to replace hours of manual library curation with one-click batch processing and fast editing workflow, and improves continuously through shared user corrections that refine the AI over time.

Description

AudioCrucible solves a universal pain point for sound designers: naming and organizing sound effect libraries. The Universal Category System (UCS) is the industry standard taxonomy, but manually classifying thousands of files against its 750+ categories is tedious, error-prone work that can consume weeks.

The application analyzes audio files using Microsoft CLAP, a neural network trained on audio-text pairs that produces semantic embeddings of what a sound “means.” These embeddings are matched against pre-computed UCS text vectors to find the best category. A local LLM (Large Language Model) then generates track titles, effect names, and descriptions grounded in the classification result.

A Deep Analysis mode goes further: it scans long ambient recordings for anomalous events (animal calls, aircraft, machinery) using bioacoustic signal processing techniques, displays them as 2D time-frequency boxes on an interactive spectrogram, and lets users listen to each event isolated to its frequency bands. A built-in speech-to-text module (Whisper) enables hands-free corrections.

In addition, when users correct a classification, they can opt to share that correction along with the audio’s AI fingerprint. These contributions are collected, validated, and used to retrain the models, making the tool progressively smarter for the entire community. Every correction benefits all future users. Everything runs locally: no audio leaves the machine. Metadata is embedded directly into WAV, FLAC, OGG, and AIFF files for compatibility with SoundMiner, BaseHead, and any DAW. Available on Windows and macOS.

ISA Entrant: Samuel Monceau