Neural Pitch Encoder Training With Frequency-Shifted Audio

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing pitch detection methods struggle with accuracy in noisy environments and require extensive annotated data for supervised learning, especially for high temporal and frequency resolution.

Innovation Solution

A self-supervised technique for training a neural network encoder to predict pitch using frequency-shifted audio samples, allowing for relative pitch estimation and calibration to determine absolute pitch values.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If supervised machine learning techniques are used to train an encoder for pitch prediction, then pitch prediction accuracy can be improved, but the cost and difficulty of generating sufficient annotated training data increases significantly

Engineering Contradiction:
Improvepitch prediction accuracyVSAvoidamount of annotated training data
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The system uses self-supervised learning where the model generates its own training signals by predicting pitch from audio samples and using those predictions to create frequency-shifted training samples. This eliminates the need for extensive manually annotated data while maintaining high pitch prediction accuracy through autonomous data generation and model refinement.

Inventive Principle:
Principle #25Self-service

2Measurement precision

If high temporal and frequency resolution is desired for pitch detection, then pitch prediction accuracy improves, but the requirement for annotated training data increases

Engineering Contradiction:
Improvetemporal and frequency resolutionVSAvoiddifficulty of generating training data
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

The system performs preliminary pitch predictions on audio samples before formal training begins. These preliminary predictions are then used to generate frequency-shifted training samples that are tailored to the specific characteristics of the audio data, enabling high-resolution pitch detection without requiring pre-prepared annotated datasets at the desired resolution.

Inventive Principle:
Principle #10Preliminary action

3Ease of operation

If heuristic algorithms are used for pitch detection, then the method is simple to implement, but performance deteriorates in the presence of noise or multiple pitch sources

Engineering Contradiction:
Improvesimplicity of implementationVSAvoidperformance in noisy environments
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The system replaces traditional heuristic algorithms with a neural network-based encoder that learns pitch detection patterns directly from data. This substitution enables the model to handle noisy environments and multiple pitch sources effectively while maintaining implementation simplicity through automated training and inference processes.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4026124B1Self-supervised pitch estimation
Publication Date: 2025.12.10 GOOGLE LLC
  • EP4026124B1 patent drawingFigure 1
  • EP4026124B1 patent drawingFigure 2
  • EP4026124B1 patent drawingFigure 3A

AI summary

Example embodiments relate to techniques for training artificial neural networks or oilier machine-learning encoders to accurately predict the pitch of input audio samples in a semitone or otherwise logarithmically-scaled pitch space. An example method may include generating, from a sample of audio data, two training samples by applying two di fferent pitch shifts to the sample of audio training data. This can be done by converting the sample of audio data into the frequency domain and then shifting the transformed data. These known shifts are then compared to the predicted pitches generated by applying the two training samples to the encoder. The encoder is then updated based on the comparison, such that the relative pitch output by the encoder is improved with respect to accuracy. One or more audio samples, labeled with absolute pitch values, can then be used to calibrate the relative pitch values generated by the trained encoder.