Music Analysis
Technology

Float-Precision Tempo. Every Beat Timestamped.

Brizm 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 picks the most likely beat sequence across the whole track, rather than greedy peak-picking. You get BPM to one decimal place, a half/double-tempo alternative, and a timestamp for every beat. It is built to cope with syncopation, tempo changes and tracks without a clear downbeat.

Pitch-Class Spectral Analysis. Benchmarked on GiantSteps.

A custom pitch-class spectrogram feeds a convolutional network trained on pitch-class spectrograms, which picks one of 24 keys (12 major, 12 minor). The spectral front end is built in-house for key detection.

On the GiantSteps key benchmark (604 electronic tracks), the network's published result is 66.7% exact and 85% exact or harmonically related (fifth, relative or parallel key). The same evaluation scores rekordbox 7 at 56.8% exact and Mixed In Key 8.3 at 69.4%. The Camelot code is derived from the key.

Six Features. Computed from Audio. Not Metadata.

A transformer-based audio embedding model, pre-trained by its authors on 3.3 million tracks labelled with music styles, turns a 30-second clip into an embedding. A small regression network, trained by Brizm on about 7,000 clips labelled with Spotify's audio-feature values, maps it to six continuous features.

Energy, danceability, happiness, acousticness, instrumentalness and speechiness are each computed from the audio, not looked up in a database. On 1,500 held-out clips by artists absent from training, they track Spotify's values at Pearson r 0.78–0.88.

Brizm Knows Where the Chorus Is.

A self-similarity analysis across the full track finds what repeats, and a novelty-based boundary detector finds where sections change. Each segment is then labelled by harmonic similarity: intro, verse, chorus, bridge or outro.

Every section has start and end times and a confidence score. Rhythm games, VJ tools and DJ software use this kind of data; new apps can no longer get it from Spotify, which closed its /audio-analysis endpoint to new apps in November 2024.

Two Models. Four Stems. GPU-Accelerated.

An ensemble of transformer and hybrid architectures powers Brizm's separation pipeline. Two-stem vocal isolation takes about 30 seconds; full 4-stem separation (vocals, drums, bass, other) takes about 45 seconds.

Both run on GPU containers that scale automatically. Results are kept for 24 hours (30 days with an active subscription) for download, then deleted; uploads are deleted as soon as processing finishes.

9.3M Tracks Resolved. Growing with Every Query.

Brizm keeps a growing catalog of acoustic data in three tiers, each serving a different layer of the stack. A cache miss triggers analysis, and the result is cached.

WASM-Accelerated Pipeline. Real Transition Forensics.

Brizm's mix analyser runs a multi-stage WASM-accelerated pipeline on the audio. Several independent detectors vote to locate each transition to sub-second precision. Phase-locked beat analysis then measures the drift, in milliseconds, at each one.

Genre profiles, each with its own tolerances and scoring weights, produce a composite grade from A to F covering technical precision, harmonic compatibility, energy flow and EQ, with the full breakdown shown.

Every score comes from the waveform, not from metadata lookups. Every transition is timestamped, and every grade breaks down into weighted sub-scores. The same file gives the same analysis.

Purpose-Built for Specific Problems.

Four tools for specific problems, each built on custom spectral analysis, WASM-accelerated pipelines and in-house detection algorithms. All four run on your device.

34 Audio Tools. Local by Default. No Sign-Up.

The tools run directly on your device — custom WASM DSP pipelines compiled from AssemblyScript, on-device neural inference, and the Web Audio API for real-time signal routing. SharedArrayBuffer lets the work run on several threads at once.

Local processing never uploads your file, needs no account and has no processing caps. Cloud Assist is optional and uploads the file. From BPM and key detection to loudness metering and chord recognition, the pipeline runs locally on your audio.

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 100 ms; a cache miss is computed on request and returns in 1–5 seconds.

Universal track resolution maps a platform ID to its known cross-references and acoustic features in one request. Also: beat grids with float-precision timestamps, audio embeddings for custom ML, song structure with section labels, and DJ mix intelligence built on co-occurrence data from 652K+ DJ set tracklists.