Constant-Q Transform (CQT)

Unlike the Short-Time Fourier Transform (STFT) which uses linearly spaced bins, the CQT spaces its frequency bins geometrically. The center frequency $f_k$ for the $k$-th bin is defined as $f_k = f_{min} \cdot 2^{k/B}$, where $B$ is the number of bins per octave and $f_{min}$ is the minimum frequency. This logarithmic spacing directly maps to musical pitch intervals, making the CQT the standard front-end for music information retrieval and harmonic analysis. Each bin corresponds to a semitone or a fraction thereof, ensuring that harmonic overtones align consistently across the representation.

The defining property of the CQT is the constant ratio of center frequency to bandwidth, known as the Q-factor. Mathematically, $Q = f_k / \Delta f_k$, where $\Delta f_k$ is the bandwidth. This is achieved by varying the window length $N_k$ inversely with frequency: $N_k = Q \cdot f_s / f_k$, where $f_s$ is the sampling rate. A high Q-factor provides fine spectral resolution for analyzing closely spaced low-frequency components, while a low Q-factor captures rapid transients at high frequencies. This adaptive resolution overcomes the fixed Heisenberg-Gabor uncertainty principle trade-off inherent in the STFT.

To maintain the constant Q-factor, the CQT employs a frequency-dependent window length. Low-frequency bins use long time windows to integrate over many cycles, achieving narrow bandwidths. High-frequency bins use short windows to resolve fast temporal events. This is the inverse of the Continuous Wavelet Transform (CWT) scaling principle, where the time support of the analyzing wavelet changes with scale. The window function $w[n, k]$ is typically a Hann or Hamming window scaled to the required length, ensuring smooth spectral leakage characteristics across all bins.

Direct computation of the CQT via time-domain convolution is computationally expensive due to the varying window lengths. The standard efficient algorithm uses a sparse kernel matrix in the frequency domain. The signal is transformed via a single large FFT, and the CQT bins are computed by multiplying the spectrum with a sparse set of pre-computed kernel coefficients. This FFT-based fast CQT avoids redundant calculations and leverages optimized FFT libraries. The kernel matrix is highly sparse because each CQT bin only interacts with a narrow band of the signal's spectrum, enabling real-time processing on modern hardware.

A well-designed CQT with sufficient bins per octave and proper window overlap is approximately invertible. The inverse CQT reconstructs the time-domain signal from the complex time-frequency coefficients. Perfect reconstruction requires the analysis windows to satisfy a non-stationary Gabor frame condition. This invertibility is critical for applications like source separation and audio effects processing, where modifications are made in the CQT domain and must be rendered back to a coherent waveform without phase artifacts. Iterative Griffin-Lim algorithms are often used to recover phase information when only magnitude coefficients are available.

The CQT organizes the spectrum into octave bands, where each octave contains exactly $B$ frequency bins. This structure naturally aligns with fractional octave analysis standards like 1/3-octave and 1/12-octave band filters used in acoustics and vibration engineering. For a 12-bin-per-octave CQT, each bin corresponds to a musical semitone, enabling direct harmonic analysis. The octave structure also facilitates multi-rate processing: higher octaves can be downsampled before analysis, reducing computational load while preserving the constant-Q property across the full spectrum.

Two fundamental visualization outputs that highlight the CQT's distinct approach:

• Spectrogram: The squared magnitude of the STFT, displaying energy on a linear frequency axis with uniform time-frequency cells. Resolution is fixed across all frequencies.
• Scalogram: The squared magnitude of the CWT coefficients, plotted on a logarithmic frequency axis. The CQT produces a similar log-frequency display where each octave contains the same number of bins, mirroring the human auditory system's perceptual pitch scaling. This makes CQT spectrograms the standard for music information retrieval tasks like chord recognition and onset detection.