ANN Magnitude Phase Representation for Transposition Invariant Music
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Music information retrieval (MIR) tasks face challenges due to high variance in musical dimensions such as transposition and tempo, leading to inefficiencies in classification and processing, as existing methods struggle to recognize similarities between shifted or tempo-variant audio signals.
Innovation Solution
A computer-implemented method for training an artificial neural network (ANN) that generates magnitude and phase representations of input signals, with a constraint to make these representations more similar despite transformations, using an auto-encoder with encoding and decoding layers to isolate transformation-dependent and independent components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional windowing methods are used to learn representations from music signals, then local patterns can be detected, but representations become highly sensitive to absolute time shifts and tempo variations, causing neighbouring representations to differ substantially even for overlapping windows
Solution Approach 1:
The patent segments the music signal into overlapping windows but then applies a transformation to make representations invariant to time shifts within those windows. By using phase difference calculations between adjacent windows, the method captures local patterns while eliminating sensitivity to absolute positioning, thus resolving the contradiction between detecting local patterns and maintaining representation stability.
Solution Approach 2:
The patent transitions from representing signals in the time domain to the frequency domain using Fourier transforms, and further to a phase difference domain. This dimensional transformation allows the system to capture temporal patterns while being invariant to absolute time shifts, as the phase difference between adjacent frequency bins encodes temporal information without depending on absolute timing.
2Measurement precision
If models are trained to recognize musical patterns, then classification accuracy improves, but the high variance in musical dimensions like transposition and tempo requires larger models with more memory consumption
Solution Approach 1:
The patent changes the representation parameters from raw time-domain signals to phase difference representations in the frequency domain. This parameter transformation inherently captures tempo-invariant and transposition-invariant features, allowing smaller models to achieve the same classification accuracy by working with more informative and invariant features rather than requiring larger capacity to handle variance.
Solution Approach 2:
The patent extracts and removes the variance components related to transposition and tempo by calculating phase differences between adjacent frequency bins. This extraction isolates the essential musical pattern information from the distracting variations, allowing models to focus on invariant features and reducing the computational resources needed for accurate classification.
3Productivity
If window size and overlap are increased to improve pattern detection, then more musical patterns can be captured, but processing time increases and generalization to shifted patterns deteriorates
Solution Approach 1:
By transforming to the frequency domain and using phase difference representations, the patent enables efficient processing of tempo variations without requiring extensive overlapping windows. The phase difference calculation inherently handles tempo scaling, allowing smaller windows with less overlap to achieve the same pattern detection capability, thus reducing processing time while maintaining detection effectiveness.
Data Source
AI summary
A computer-implemented method of training an artificial neural network (ANN) by generating one or more learned parameters for use during a subsequent inference phase of the trained ANN, comprises providing training data representing first and second input signals, the second input signal exhibiting one or more transformations relative to the first signal selected from a set of transformations; using the ANN and in response to the one or more parameters, generating a magnitude and phase representation of each of the first and second input signals; and training the one or more parameters, in dependence upon a constraint which causes the magnitude representation of the first input signal and the magnitude representation of the second input signal to tend to become more similar to one another, the training step comprising: detecting an error signal; and updating the one or more parameters in dependence upon the error signal.