Pitch Estimation with CREPE

The task of estimating the fundamental frequency of a monophonic sound recording, also known as pitch tracking, is fundamental to audio processing with multiple applications in speech processing and music information retrieval.

CREPE (Convolutional REpresentation for Pitch Estimation) is a deep learning-based pitch estimation algorithm that was proposed in 2018. It is a convolutional neural network (CNN) that operates directly on the time-domain waveform of an audio signal to estimate its pitch.  The CNN learns to extract features from the time-domain waveform that are relevant to pitch estimation. These features are then used to predict the pitch of the audio signal.

CREPE is trained on a large dataset of audio signals with known pitch labels. The training dataset includes a variety of audio signals, such as speech, music, and environmental noise. This allows CREPE to learn to estimate the pitch of a wide variety of audio signals.

Spectral Centroid

The spectral centroid is a fundamental audio feature that characterizes the "center of mass" or the average frequency of a sound spectrum. It plays a crucial role in audio analysis, including pitch estimation and timbre characterization.

• A higher spectral centroid indicates that the audio signal contains a significant amount of high-frequency content.

• Conversely, a lower spectral centroid suggests that the audio signal is dominated by low-frequency components.

• It provides insights into the "brightness" or "dullness" of a sound, with higher values associated with brighter sounds and lower values with duller sounds.

Here's a brief comparison between a few common methods for pitch estimation:

Autocorrelation:

• Autocorrelation measures the similarity between a signal and a time-shifted version of itself. Peaks in the autocorrelation function correspond to periods or pitches.

• Good for periodic signals with clear fundamental frequencies, but susceptible to noise interference.

• Simple to implement algorithmically, involves calculating correlations at different time shifts.

• Widely used in speech analysis due to its simplicity and effectiveness in detecting fundamental frequencies in voiced speech.

Fast Fourier Transform (FFT):

• FFT analyzes the frequency content of a signal by transforming it from the time domain to the frequency domain.

• Effective for signals with clear and steady tones but less accurate with noisy or transient signals.

• Efficient computationally, especially when using optimized FFT algorithms.

• Commonly used in music applications for pitch analysis and detecting fundamental frequencies in harmonic-rich signals.

Harmonic Product Spectrum (HPS):

• HPS involves combining multiple spectra (by multiplying them) to enhance the harmonic structure in a signal.

• Effective for identifying harmonic structures, especially in music signals.

• Moderately complex due to spectral analysis and combining spectra.

• Useful in music analysis where harmonic content and identifying fundamental frequencies among harmonics are crucial.

YIN Algorithm:

• YIN algorithm calculates the difference between a signal and its shifted version to estimate the fundamental frequency.

• Provides high precision and robustness, even in noisy environments.

• Moderately complex due to advanced algorithms, involving difference calculations and thresholding.

• Commonly used in both music and speech analysis due to its accuracy in detecting pitch, especially in challenging conditions.

Subharmonic Summation:

• This technique involves detecting and summing subharmonic frequencies (integer fractions of the fundamental frequency).

• Effective for signals with strong subharmonic content, aiding in identifying the fundamental frequency.

• Can be complex, especially when dealing with multiple subharmonics or non-linear systems.

• Useful in specific musical contexts and specialized applications where subharmonic relationships are prominent, like some wind instruments or vocal techniques.