Skip to content
Multigrid

Understanding audio that is not speech

Tagging, embedding and anomaly detection over sound — music, machinery, wildlife and everything that is not a transcript.

Speech recognition has absorbed most of the attention paid to audio machine learning, and it has left an odd gap: the great majority of sound in the world is not somebody talking. A bearing beginning to fail, a warbler at four in the morning, a snare hit, a cough, a breaking window — none of these has a transcript, and a model that turns audio into words is the wrong tool for all of them. What these problems share is a different pipeline: turn the waveform into a time–frequency image, run something over that image, and get back labels, a score, or a vector.

These pages work from that pipeline outward. The transform’s parameters — sample rate, window length, hop length, how many mel bands and where they start — are treated as decisions with consequences rather than as boilerplate, because they are where most of the information is lost and most of the accidental failures begin. Where a page names a model, a dataset or a metric, it names the real one and says what the published documentation actually claims.

Music Genre Classification: How a Model Hears Genre

The spectrogram features a genre classifier actually consumes, assembled into one worked feature vector, and why the label is the weakest part of the system.

9 min read

Audio Tagging Models: Multi-Label Sound Classification Explained

Why tagging uses independent sigmoids rather than a softmax, how frame scores become clip tags, and what the published mAP figures do and do not tell you.

9 min read

Identifying Bird Species From Audio Recordings

How a bird identifier turns three seconds of dawn chorus into a species score, and the specific reason overlapping calls degrade it.

9 min read

Music Recommendation From Audio Embeddings, Not Metadata

Where collaborative filtering structurally cannot help you, and what a content-based audio embedding returns instead, with the nearest-neighbour arithmetic worked out.

9 min read

Acoustic Anomaly Detection for Industrial Machinery

A worked FFT baseline-versus-deviation calculation on a machine spectrum, and why the naive version of it produces alarms nobody trusts.

10 min read

Bioacoustic Monitoring: Detecting Wildlife From Continuous Audio

The storage and compute cost of running one autonomous recorder for a season, derived from stated assumptions, and what the arithmetic changes about the design.

10 min read

Classifying Music Mood and Tempo Automatically

How beat tracking gets a tempo out of an onset envelope, why octave errors are endemic, and what the valence/arousal plane does to a mood label.

9 min read

Audio Embeddings for Similarity Search Across a Sound Library

How frame-level audio models become one vector per clip, and the evaluation protocol that tells you which embedding is better for your library.

10 min read

Instrument Recognition From a Polyphonic Recording

Why a softmax classifier cannot survive two instruments playing at once, and exactly what multi-label training with partial labels changes.

9 min read

Classifying Cough and Respiratory Sounds From Audio

How cough and breath sound models are built and evaluated, why shortcut learning is endemic in this literature, and what the false-positive arithmetic looks like at realistic prevalence.

10 min read

Detecting Gunshots From Acoustic Sensor Networks

How a network of microphones turns three arrival times into a map coordinate, and where that arithmetic goes wrong.

11 min read

Detecting Mechanical Faults From Engine Sound

How rotating machinery encodes its own faults at predictable frequencies, and how to compare a recording against that signature.

11 min read

Audio Embedding Models Like CLAP, Explained

How contrastive audio-text training produces a shared vector space, and why that lets a typed phrase retrieve a sound.

11 min read

Underwater Acoustic Monitoring With AI

What a hydrophone deployment actually produces, derived from its sample rate, and what the recovery cycle does to your modelling choices.

11 min read

Detecting a Baby Crying or a Household Sound Event at Home

Why a 99%-specific detector still fires hundreds of times a day at home, and what actually fixes it.

10 min read

Classifying Speaker Emotion From Non-Verbal Vocal Cues

Which acoustic parameters carry affect, what they reliably distinguish, and where the transcript is the better signal.

10 min read

What Continuous Audio Monitoring Costs to Run All Day

A full derivation of the daily and monthly cost of listening to one audio stream continuously, from sample rate and window size upward.

11 min read

Detecting Applause, Laughter and Crowd Reaction in Recordings

What separates applause from laughter acoustically, and a worked precision and recall example for the case where it does not.

10 min read

Music Key and Chord Detection From Audio

How folding a spectrum into twelve pitch classes produces the representation both key and chord estimation are built on.

11 min read

A Classification Model Suddenly Started Getting Audio Wrong

A sudden step drop in audio classifier accuracy usually means clipping upstream; here is the check that finds it and the fix that works.

10 min read

Other topics