How does Shazam work? (2022) (cameronmacleod.com)

Article summary

The article explains how Shazam works by using audio fingerprinting to identify songs. It calculates a spectrogram of the audio, finds peaks in the spectrogram, hashes these peaks, and then matches them to a database of known songs. The algorithm is designed to be robust to noise and distortions, allowing it to identify songs even in noisy environments. The article also discusses the history of Shazam and its development.

Main themes

• audio fingerprinting
• Shazam algorithm
• machine learning for audio identification
• noise robustness
• hash-based vs model-based approaches
• music information retrieval
• audio recognition technology
• data augmentation and denoising techniques

What commenters say

• Some commenters argue that using hashes is still a superior approach to using machine learning models for audio identification due to its efficiency and robustness.
• Others suggest that a modern approach using embeddings and contrastive learning could be effective for audio identification.
• There is disagreement about whether adding noise to the training set or attempting to denoise the input is a better approach for training a machine learning model for audio identification.
• Some commenters believe that using a model to deconstruct a song into features like key, tempo, and chord progressions could enable more robust audio identification.
• It is argued that Shazam's success was not just due to its core recognition technology, but also its business processes and supporting systems.
• There is a discussion about the trade-offs between using a hash-based approach versus a model-based approach for audio identification, with some arguing that the former is more suitable for large-scale audio identification.
• The use of denoising algorithms versus data augmentation techniques for improving the robustness of audio identification models is also debated.
• Some commenters note that the problem of audio identification is distinct from other recognition tasks, such as face recognition, and may require specialized approaches.