audio-ml

A desktop app to fine-tune, generate and perform live music with diffusion models. Runs on MacOS, Windows and Linux.

Code to train a custom time-domain autoencoder to dereverb audio

Dual-model speech AI toolkit for speaker verification and speaker-aware diarization, with streaming inference, meeting analysis, long-audio monitoring, and speaker-bank integration.

Local MCP server + CLI turning YouTube & local audio into rich sonic signatures. Extracts BPM, section-by-section key, vocal presence, transient punch, and 512-dim CLAP vibe embeddings. Powered by Demucs stem separation & librosa. 100% private, offline-first, and GPU-accelerated with graceful CPU/HPSS degradation.

From-scratch audio-AI research system + 99-lesson bilingual course — DSP · ASR · Music · LLM/RAG, every core hand-written, no API wrappers.

Unified dual-teacher distillation (ReDimNet + ASSIST) into Wav2Vec2 for speaker verification and deepfake detection

Key Features: Simple VAE architecture with encoder/decoder Synthetic music data generation for training Interactive training with progress tracking Music generation from latent space sampling Audio conversion and playback Downloadable audio files

An intelligent, automated video switcher for live performances

This repository consists of projects where GANs were used to work with audio data like music generator etc.

Real-time speech enhancement pipeline — custom-trained U-Net denoising model, ONNX inference, Overlap-Add synthesis, and virtual audio routing for Teams, Zoom, and DAW use. CPU-only, no cloud dependency.

Stable Audio LoRA Trainer of salty goodness

Researching compact, feedback-driven music decision models for preference learning, creative-intent evaluation, and integration with larger AI systems.

Audio analysis in javascript/typescript

Engine identification using acoustic signal analysis and machine learning to classify 8 vehicle types. Audio signals are processed using FFT and feature extraction, and a multi-class model predicts vehicle categories based on their unique sound patterns.

Recognizing material families from sliding audio and vibration, with a focus on generalization across surfaces, speeds, and sensor setups.

Edge-deployable keyword spotter: INT8-quantized DS-CNN on Google Speech Commands, exported to ONNX, with fp32 vs INT8 benchmarks, a live mic demo, and a C++ inference harness.

Convert Meta's HTDemucs (Hybrid Transformer Demucs) to Apple Core ML. Real-valued STFT/ISTFT wrapper, manual MHA decomposition, pre-computed overlap-add. Includes Swift example.

Attention routing for F1 race radio — surfaces the moments in a session where someone should have listened. Arousal x valence over real team radio, joined to official lap telemetry.