Music Analysis
Technology

Float-Precision Tempo. Not a Rounded Integer.

TuneLab feeds audio through a multi-resolution spectral analysis stage, then into a recurrent neural architecture with long-term temporal context that outputs beat and downbeat probabilities across the full track.

A probabilistic sequence model then finds the globally optimal beat sequence — no greedy peak-picking, no dropped beats. The result: float-precision BPM (±0.1), half/double-tempo alternatives, and per-beat timestamps. Handles syncopation, polyrhythms, tempo changes, and even tracks with no clear downbeat.

Pitch-Class Spectral Analysis. 82.1% on GiantSteps.

A custom pitch-class spectrogram feeds a convolutional network trained on pitch-class spectrograms for 24-class key classification (12 major + 12 minor). No external signal processing libraries — the entire spectral frontend is purpose-built for key detection.

Tested against the GiantSteps-MTG benchmark (the standard dataset for key detection evaluation), TuneLab achieves 82.1% accuracy — significantly outperforming classical template-matching approaches. Camelot wheel mapping is computed automatically from the 24-class output.

Six Features. Computed from Audio. Not Metadata.

A transformer-based audio embedding model, pre-trained on 100K+ annotated tracks, produces a rich spectral representation from a 30-second clip. Specialised regression heads fine-tuned on curated reference data then extract six continuous features.

Energy, danceability, happiness, acousticness, instrumentalness, speechiness — each genuinely computed from the audio waveform, not looked up in a database. Holdout metrics are published because TuneLab's numbers hold up to scrutiny.

TuneLab Knows Where the Drop Is.

A self-similarity analysis across the full track detects structural repetition, then a novelty-based boundary detector locates transitions between sections. Hierarchical clustering labels each segment: intro, verse, chorus, drop, breakdown, outro.

Every section comes with start/end timestamps and a confidence score — the kind of structural data that rhythm games, VJ tools, and DJ software need but lost when Spotify deprecated their /audio-analysis endpoint.

Two Models. Four Stems. GPU-Accelerated.

An ensemble of state-of-the-art transformer and hybrid architectures powers TuneLab's separation pipeline. Two-stem vocal isolation produces studio-quality splits in roughly 30 seconds; full 4-stem separation (vocals, drums, bass, other) completes in approximately 45 seconds.

Both run on dedicated GPU containers with automatic scaling, custom-optimised for high-throughput inference. Results are stored as presigned URLs with 24-hour expiry. No audio is retained beyond that window.

9.3M Tracks Resolved. Growing with Every Query.

TuneLab maintains a continuously growing catalog of acoustic data across three tiers — each serving a different layer of the intelligence stack. Cache misses trigger real-time DSP and are cached permanently.

Live Beatmatching. Two Streams. 10ms Tick.

A digital phase-locked loop continuously tracks tempo and phase across two live audio streams, correcting pitch in real time to keep them beatmatched. The control loop runs at a 10ms tick rate with ±2% pitch bend range — fast enough to lock onto tempo drift within seconds.

This is the engine behind the live radio demo — two independent internet radio streams, beatmatched automatically, mixed live in the browser. No pre-analysis, no metadata. Just the audio signal.

WASM-Accelerated Pipeline. Real Transition Forensics.

TuneLab's mix analyser runs a multi-stage WASM-accelerated pipeline on actual audio: multiple independent detection methods work in concert through a voting system to locate transitions with sub-second precision. Phase-locked beat analysis then measures drift in milliseconds at each one.

Multiple genre profiles — each with calibrated tolerances and scoring weights — produce a composite assessment: technical precision, harmonic compatibility, energy flow, and EQ quality, graded A to F with a fully transparent breakdown.

Every score is derived from the waveform itself — not from metadata lookups or pre-computed averages. Every transition is timestamped. Every grade is decomposed into sub-scores with transparent weighting. The full methodology is published, and results are reproducible: same audio in, same analysis out.

Purpose-Built for Problems Others Ignore.

Four tools that solve problems no one else addresses — each built on custom spectral analysis, WASM-accelerated pipelines, and proprietary detection algorithms. All processing runs on your device.

32 Audio Tools. Zero Uploads. No Limits.

Every tool runs directly on your device — custom WASM DSP pipelines compiled from AssemblyScript, ONNX Runtime for neural inference, and the Web Audio API for real-time signal routing. SharedArrayBuffer enables true multi-threaded processing with near-native throughput.

No file ever leaves your machine. No account required. No processing caps. From BPM detection and key analysis to loudness metering and chord recognition — the full pipeline executes locally, with results computed from your actual audio waveform.

Real Analysis. Not Cached Metadata.

A REST API built on the same DSP and neural inference that powers the tools. Synchronous responses for lookups and analysis — no polling loops, no per-request surprises. Cache hits return in under 100ms; cache misses trigger real-time compute and respond in 1–5 seconds.

Universal track resolution maps any platform ID to all known cross-references plus verified acoustic features — one request, every major streaming service resolved. Beat grids with float-precision timestamps, audio embeddings for custom ML, song structure with section labels, and DJ mix intelligence combining co-occurrence data from 652K+ analysed sets.

The Engine Powers Working Products.

AI DJ Radio — running since 2020. 24/7/365 continuous mix of underground electronic music, with automated track selection driven by energy curves, harmonic compatibility, and genre coherence. No human intervention. No pre-programmed playlists. The engine described on this page was built to make this work.

Coming soon — a real-time audio platform. Powered by TuneLab.