A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Music Information Retrieval (MIR) is the research field and engineering discipline concerned with extracting musically meaningful information from audio signals — automatically and computationally. MIR enables technologies like music recommendation (Spotify's Discover Weekly), automatic chord recognition, tempo detection, key estimation, music transcription, cover song identification, and genre classification.
MIR begins with feature extraction — computing numerical descriptors from audio that capture musically relevant properties. Temporal features describe signal energy and changes over time: RMS energy, zero-crossing rate, onset detection (finding the start of new notes or percussion hits). Spectral features describe the frequency content: spectral centroid (perceived brightness), spectral rolloff, spectral flux (rate of change between frames). Mel-frequency cepstral coefficients (MFCCs) are derived from a perceptually-scaled frequency representation (the Mel scale) and computed via a cepstral transformation — they efficiently encode the timbre and vocal/instrumental quality of audio and are the most widely used features in speech and music recognition.
Pitch estimation and chord recognition require identifying the fundamental frequency of pitched sounds from polyphonic audio. Constant-Q Transform (CQT) provides a frequency representation with logarithmic frequency resolution that aligns with musical pitch spacing (each octave spanning the same number of bins, matching how humans perceive pitch). Chromagram representations fold pitch into pitch class (the 12 notes of the chromatic scale, discarding octave information), enabling key and chord analysis.
Beat tracking and tempo estimation use onset detection and autocorrelation to find the pulse underlying rhythmic audio. Dynamic time warping (DTW) aligns two audio sequences of different tempos or durations, enabling score-to-audio alignment, cover song detection, and performance comparison.
Music Information Retrieval sits at the intersection of signal processing, machine learning, and musicology. Early MIR research (1990s–2000s) focused on handcrafted features and classical machine learning (SVM, k-NN, random forests). Modern MIR is dominated by deep learning: convolutional neural networks applied to mel spectrograms learn to classify genre, detect chords, or transcribe music directly from learned feature representations rather than handcrafted ones.
The applications of MIR are ubiquitous in the music industry. Streaming platforms use audio fingerprinting (based on spectral peak matching) for copyright identification (Shazam's algorithm, Gracenote). Recommendation systems combine audio feature similarity with collaborative filtering (user behavior). Automatic mixing systems (AI mastering services) use MIR to analyze tracks and apply genre-appropriate processing. Music education apps (Yousician, Simply Piano) use pitch detection to give real-time feedback on performance.
Research challenges in MIR include: polyphonic transcription (converting audio with multiple simultaneous notes into symbolic notation), lyrics alignment (finding where each sung word occurs in the audio), musical genre classification (inherently subjective and culture-dependent), and source separation (isolating individual instruments from a mixed recording). Demucs and Spleeter demonstrate recent deep learning progress on source separation, with commercially deployed applications in stem extraction services used by DJs, producers, and remix artists.
Topics in reflective domains aren't scored by quiz answers. Read, reflect, and mark when you've thought it through.