Abstract
This paper presents a framework for robust music identification and detection by treating audio as a sequence of "music phones," analogous to speech recognition. Using Weighted Finite-State Transducers (WFSTs) and Gaussian Mixture Models (GMMs), the system achieves high accuracy on short, noisy snippets. The approach also incorporates a Universal Background Model (UBM) and Support Vector Machines (SVM) for reliable "out-of-set" song detection.
Introduction
Traditional music identification systems typically rely on "audio hashing," which creates digital fingerprints for exact matching. While efficient, these methods often struggle with noise, distortions, and determining if a query is present in the database at all. This research addresses these challenges by modeling music as a language, providing robustness against background noise, speed changes, and low-bitrate compression.
Methods
The core of the system is the use of Weighted Finite-State Transducers (WFSTs). The audio is first decomposed into "music phones"—atomic sound units learned through unsupervised clustering of Mel-frequency cepstral coefficients (MFCCs). A recognition transducer is then constructed to map these phone sequences to specific song identifiers.
To handle real-world audio degradation, the system employs an "edit distance transducer" that allows for fuzzy matching by accounting for insertions, deletions, or substitutions of phones. For "out-of-set" detection, the researchers utilize a Universal Background Model (UBM) to represent the average sound of all music. An SVM classifier then evaluates the likelihood ratio between the specific song model and the UBM to decide if a snippet belongs to the database.
Results and Analysis
The system was evaluated on a dataset of over 15,000 songs with various distortions. A key insight from the analysis is "factor uniqueness": as the length of a phone sequence increases, the probability of it appearing in multiple songs drops sharply. By 10 seconds, sequences are almost entirely unique.
As shown in the results, the system achieved 99.4% identification accuracy on clean data and remained highly resilient to MP3 compression (95.5% at 32 kbps). However, performance degrades under extreme noise (11.1% at 5.9 dB SNR) and significant speed shifts (+/- 10%), which cause spectral shifts the models were not trained to handle.
Conclusion
By applying speech recognition paradigms to music, this framework provides a robust alternative to traditional hashing. Future work aims to improve invariance to pitch and speed changes through time-scaling methods or augmented training data, and to explore the use of FSTs for structural music analysis.
Summary updated