Fact-checked by Grok 2 weeks ago

Short-time Fourier transform

The short-time Fourier transform (STFT) is a fundamental tool in that computes the of short, overlapping segments of a signal, obtained by multiplying the signal with a sliding , to reveal how its content evolves over time. This approach addresses the limitations of the standard , which assumes stationarity and loses temporal information for non-stationary signals like speech or . By localizing the analysis in both time and domains, the STFT produces a two-dimensional representation known as a when taking the magnitude squared, enabling the identification of transient events and time-varying spectral features. Introduced by Hungarian-British physicist Dennis Gabor in 1946 as part of his work on communication theory, the STFT—originally termed the "Gabor transform"—marked an early effort to adapt Fourier analysis for practical applications in quantum mechanics and signal transmission. Gabor's innovation involved windowing the signal with a Gaussian function to balance time and frequency localization, laying the groundwork for modern time-frequency methods despite the computational challenges of the era. Subsequent developments in the 1970s and 1980s, including efficient discrete implementations, expanded its use in digital signal processing, particularly with the advent of fast Fourier transform algorithms. Mathematically, the continuous STFT of a signal f(t) with respect to a window function g(t) is given by
\mathcal{V}_g f(\tau, \omega) = \int_{-\infty}^{\infty} f(t) \overline{g(t - \tau)} e^{-2\pi i \omega t} \, dt,
where \tau denotes time location and \omega frequency, providing the inner product between the signal and modulated, shifted versions of the window. In the discrete case, for a sampled signal x and window w of length M, the STFT at time index m and frequency \omega is
X_m(\omega) = \sum_{n=-\infty}^{\infty} x(n) w(n - mR) e^{-j \omega n},
with R as the hop size controlling overlap; this is typically computed via the discrete Fourier transform on each windowed segment. The choice of window (e.g., Hamming or rectangular) influences resolution and sidelobe suppression, while the constant overlap-add (COLA) property ensures perfect reconstruction of the original signal under certain conditions.
The STFT finds wide applications in audio and speech processing for tasks such as pitch detection and noise reduction, as well as in radar, seismology, and biomedical signal analysis like EEG for capturing dynamic frequency shifts. However, its fixed window size imposes a fundamental trade-off: shorter windows yield better time resolution but poorer frequency resolution, and vice versa, as dictated by the Heisenberg uncertainty principle in signal processing. This limitation has spurred alternatives like wavelet transforms, though the STFT remains a cornerstone due to its simplicity and invertibility.

Fundamentals

Definition and Motivation

The short-time Fourier transform (STFT) emerged in the 1940s as a pivotal advancement in signal analysis, pioneered by Dennis Gabor to extend the classical Fourier transform for capturing time-varying frequency content in non-stationary signals, such as speech and acoustic phenomena. Gabor's motivation stemmed from the need to balance time and frequency localization, drawing parallels to the Heisenberg uncertainty principle in quantum mechanics, where precise measurement in one domain inherently limits resolution in the other. This approach addressed the shortcomings of the global Fourier transform, which assumes signal stationarity and yields only an averaged frequency spectrum without temporal context. At its core, the STFT operates by segmenting the input signal into overlapping short intervals, applying a to isolate each segment, and performing a on the windowed portion to generate a two-dimensional time-frequency representation. This localized analysis reveals how frequency components evolve over time, enabling the study of transient events in signals where spectral properties change dynamically. The general continuous-time form of the STFT for a signal x(t) and window w(t) is expressed as X(\tau, \omega) = \int_{-\infty}^{\infty} x(t) \, w(t - \tau) \, e^{-j \omega t} \, dt, where \tau denotes the time shift parameter localizing the analysis window, and \omega is the angular frequency. By providing simultaneous time and frequency resolution, the STFT surpasses the global Fourier transform's limitations for non-stationary processes, facilitating applications in fields like communications and radar that demand insight into evolving signal characteristics. Post-World War II, it gained traction in signal processing for acoustics and early speech analysis, building on Gabor's foundational ideas of "elementary signals" akin to information quanta.

Window Functions

Window functions play a crucial role in the (STFT) by tapering the signal segments to mitigate and discontinuities at the edges of finite-duration analysis frames. occurs when abrupt truncation of the signal introduces high-frequency artifacts in the , while edge effects arise from the implicit assumption of periodicity in ; windowing smoothly attenuates the signal toward zero at the boundaries, reducing these artifacts and improving the accuracy of time-frequency representations. This tapering is essential for balancing localization in time and frequency, as the modulates each segment before applying the . Several common window functions are employed in STFT applications, each offering distinct trade-offs in performance. The rectangular window applies no tapering, resulting in abrupt edges and high sidelobe levels in the frequency domain, which leads to significant leakage but provides the narrowest mainlobe for optimal frequency resolution. In contrast, the Hann and Hamming windows introduce a smooth cosine-based taper, substantially reducing sidelobe levels and leakage at the cost of a wider mainlobe; the Hann window, defined as w(n) = 0.5 \left(1 - \cos\left(\frac{2\pi n}{N-1}\right)\right) for n = 0 to N-1, achieves about -31 dB sidelobe attenuation, while the Hamming variant adjusts the weighting for slightly better far-sidelobe suppression. The Gaussian window, optimal for the Gabor transform—a special case of STFT—exhibits minimal time-frequency spread due to its bell-shaped profile and self-dual Fourier transform, minimizing uncertainty in localization without sidelobes, as originally proposed in communication theory for atomic signal analysis. The Blackman window further enhances sidelobe suppression to around -58 dB through a three-term cosine series, offering improved dynamic range for spectral analysis. Finally, the Kaiser window provides flexible control via its shape parameter \beta, approximating the optimal prolate spheroidal wave function; higher \beta values (e.g., \beta > 5) yield lower sidelobes (down to -80 dB or more) but broader mainlobes, making it adaptable for applications requiring tunable attenuation. Key properties of window functions determine their suitability for STFT, including mainlobe width, which inversely affects by dictating the smallest distinguishable frequency separation, and sidelobe levels, which quantify leakage by measuring the of secondary peaks relative to the mainlobe (typically in below the peak). The time-bandwidth product, defined as the product of the window's effective duration and , characterizes the inherent ; for instance, the Gaussian achieves the minimum value of \frac{1}{4\pi} (in continuous time), aligning with the , while other windows like the Hann approach but exceed this bound. These properties are evaluated in the via the window's , where narrower mainlobes favor precise frequency estimates but risk higher interference from sidelobes, and vice versa. The length of the window introduces a fundamental in STFT performance: shorter windows (e.g., 256 samples) enhance time by capturing rapid signal changes but degrade due to broader mainlobes, while longer windows (e.g., 2048 samples) improve discrimination at the expense of temporal smearing, particularly for non-stationary signals. This depends on the signal's characteristics, with empirical guidelines suggesting window lengths proportional to the inverse of the expected variation rate. In the discrete-time STFT, windows are often normalized to preserve energy across overlapped segments, typically by scaling such that \sum_{n=0}^{N-1} w^2(n) = 1, ensuring the total power in the reconstructed signal matches the original via when using overlap-add methods. This normalization prevents amplitude distortion in inverse transformations and maintains consistent scaling.

Forward Transform

Continuous-Time Formulation

The continuous-time short-time Fourier transform (STFT) of a signal x(t) is defined as X(\tau, \omega) = \int_{-\infty}^{\infty} x(t) \, w(t - \tau) \, e^{-j \omega t} \, dt, where \tau denotes the center time of the analysis window w(t - \tau), and \omega is the . This formulation provides a time-frequency representation by applying the to successive windowed portions of the signal. The integral represents the inner product between the signal x(t) and a modulated version of the centered at time \tau, yielding the complex amplitude of the component \omega localized around that time instant. This interpretation highlights the STFT's role in capturing how the signal's content evolves over time, with the w(t) controlling the . An alternative expression arises by first windowing the signal to form x(t) w(t - \tau) and then computing its Fourier transform evaluated at \omega. Equivalently, the STFT can be viewed as X(\tau, \omega) = \mathcal{F}\{x(t) w(t - \tau)\}(\omega), where \mathcal{F} denotes the operator. Another form emphasizes : X(\tau, \omega) = e^{-j \omega \tau} \int_{-\infty}^{\infty} x(t) \, [w(t - \tau) e^{-j \omega (t - \tau)}] \, dt, interpreting it as the inner product with a time-shifted and frequency-modulated window. This continuous formulation assumes the signal x(t) extends over infinite duration, which is ideal for theoretical but contrasts with practical finite-length signals. The w(t) is typically even for in time and and possesses compact to localize the effectively. Regarding units and scaling, the STFT maintains consistency with the standard , where time t and \tau are in seconds, and \omega is in radians per second. Energy preservation follows from applied to the underlying : for fixed \tau, \frac{1}{2\pi} \int_{-\infty}^{\infty} |X(\tau, \omega)|^2 \, d\omega = \int_{-\infty}^{\infty} |x(t) w(t - \tau)|^2 \, dt, equating the in the to that of the windowed signal segment. Over the full time-frequency plane, \frac{1}{2\pi} \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} |X(\tau, \omega)|^2 \, d\omega \, d\tau = \int_{-\infty}^{\infty} |x(t)|^2 |w(t)|^2 \, dt, demonstrating that the total is preserved, weighted by the window's ; \int_{-\infty}^{\infty} |w(t)|^2 \, dt = 1 yields exact equality to the signal's .

Discrete-Time Formulation

The discrete-time short-time Fourier transform (STFT) adapts the continuous-time for digitally sampled signals, representing them as sequences x where n is the integer sample index. This enables practical computation on finite-length data, bridging theoretical analysis with applications. The discrete STFT is defined as X(m, \Omega) = \sum_{n=-\infty}^{\infty} x \, w[n - mR] \, e^{-j \Omega n}, where X(m, \Omega) is the transform at time frame m and frequency \Omega, x is the input signal, w is the window function centered at frame m, and R is the hop size in samples denoting the frame advance. The frequency variable \Omega is typically discretized as \Omega_k = 2\pi k / N for k = 0, 1, \dots, N-1, corresponding to the bins of an N-point . Here, N is chosen as a power of 2 to leverage the efficiency of the algorithm for computing each windowed segment. The hop size R determines the temporal spacing between frames and the degree of overlap between adjacent windows, with overlap given by $1 - R/M where M is the length in samples. A 50% overlap (R = M/2) is commonly used to balance time resolution and computational load while ensuring sufficient redundancy for applications like analysis. This overlap introduces redundancy in the time-frequency representation, quantified by the factor M/R, which exceeds 1 for R < M and facilitates information preservation across without aliasing in the time domain. For finite-length signals of length L, the infinite summation in the STFT definition is handled by zero-padding the signal with zeros beyond its duration or by periodic extension, ensuring all windowed segments fit within the extended sequence. Zero-padding to a total length of at least L + (M-1)(K-1) (where K is the number of frames) accommodates edge effects without distorting the core analysis. Each padded windowed segment is then transformed via an N-point , with N \geq M to capture the full spectral content. A variant known as the sliding DFT offers an efficient mechanism to update the transform incrementally for consecutive frames, exploiting the overlap to avoid full recomputation of the DFT.

Inverse Transform

Reconstruction Principles

The short-time Fourier transform (STFT) is invertible provided the window function satisfies the partition of unity condition or generates a frame with finite bounds, where redundancy arises from overlapping time shifts that ensure complete coverage of the signal. This invertibility guarantees perfect reconstruction of the original signal from its STFT representation, as long as the window is square-integrable and non-zero almost everywhere. The continuous-time inverse STFT, under perfect reconstruction conditions, is expressed as x(t) = \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} X(\tau, \omega) \, w(t - \tau) \, e^{2\pi i \omega t} \, d\tau \, d\omega, where X(\tau, \omega) denotes the STFT, w(t) is the window function (L^2-normalized such that \int |w(t)|^2 \, dt = 1), and the integral synthesizes the signal by summing modulated and shifted window contributions weighted by the STFT values. This formula holds when the window ensures the analysis-synthesis operator is bounded and invertible on L^2(\mathbb{R}). From the perspective of frame theory, the continuous STFT generates a tight frame when the window is suitably normalized, with equal frame bounds A = B = \frac{1}{\int_{-\infty}^{\infty} |w(t)|^2 \, dt} for cases where the window supports orthogonality in the time-frequency shifts. This tight frame structure provides stable reconstruction via the dual frame, which coincides with the original window, and bounds the reconstruction error independently of the signal. Energy preservation follows from an extension of Parseval's theorem to STFT frames, stating that for a signal x \in L^2(\mathbb{R}), \|x\|^2 = A \iint |X(\tau, \omega)|^2 \, d\tau \, d\omega, where A is the lower frame bound, confirming that the transform preserves the signal's L^2 energy up to the frame scaling. This property holds for tight frames and underpins the numerical stability of STFT-based processing.

Overlap-Add and Overlap-Save Methods

The overlap-add (OLA) method is a practical discrete-time algorithm for reconstructing a signal from its short-time Fourier transform (STFT) coefficients by processing overlapping frames. In this approach, each frame of the STFT is transformed back to the time domain via an inverse discrete Fourier transform (IDFT), typically implemented using the inverse fast Fourier transform (IFFT), and then windowed again before the overlapping segments are summed to form the reconstructed signal. This method ensures efficient reconstruction when the original signal is segmented into frames using a hop size R, where each frame is of length N, and the windows satisfy the constant overlap-add (COLA) condition. Perfect reconstruction in OLA is achieved if the sum of the shifted window functions equals a constant, specifically \sum_k w(n - kR) = c for some constant c \neq 0, where w(n) is the analysis window; normalization by c yields exact recovery for unmodified STFT frames. For cases involving a synthesis window identical to the analysis window, as in weighted overlap-add (WOLA) processing, the condition strengthens to \sum_k w^2(n - kR) = constant to preserve energy and avoid amplitude distortions. Common windows like the Hann function satisfy this squared COLA condition at 50% overlap, where R = N/2, enabling perfect reconstruction without aliasing for N-point IDFTs. The overlap-save (OLS) method, in contrast, is particularly suited for STFT-based filtering or convolution tasks, where it discards the transient-affected portions of each overlapped frame to retain only the valid, non-aliased middle sections for reconstruction. In OLS, the input signal frames overlap by L-1 samples (with L as the effective filter length), the IDFT outputs are computed, and the first L-1 samples per frame—contaminated by circular convolution artifacts—are discarded, while the remaining R = N - L + 1 samples are saved and concatenated directly. This approach avoids the additive summation of OLA, making it more computationally efficient for real-time processing in STFT-domain convolutions, though it requires careful hop size selection to cover the signal without gaps. In practice, a hop size of R = N/2 (50% overlap) is widely adopted for both OLA and OLS with N-point DFTs, as it balances computational load and reconstruction quality, particularly for the Hann window, which inherently meets the COLA criteria under this overlap. The COLA condition can be verified numerically by ensuring the summed window squares yield a flat response across all time indices, preventing amplitude modulation in the reconstructed signal. For modified STFTs—such as those altered in magnitude for audio effects—numerical stability requires preserving phase information during processing and applying consistent windowing to mitigate errors from inconsistencies between modified magnitudes and original phases, often necessitating iterative refinement techniques.

Time-Frequency Properties

Resolution Trade-offs

The short-time Fourier transform (STFT) inherently involves a trade-off between time and frequency resolution, dictated by the choice of analysis window. The time resolution, denoted as Δt, is approximately equal to the duration of the window function w(t), determining how precisely transient events in the signal can be localized in time. Conversely, the frequency resolution Δf is roughly the reciprocal of this duration, approximately 1 / duration of w(t), which governs the ability to distinguish closely spaced frequency components. This inverse relationship means that improving resolution in one domain degrades it in the other, a fundamental limitation arising from the nature of applied to finite segments of the signal. This trade-off is formally captured by an analogy to the Heisenberg uncertainty principle in signal processing, which states that a signal cannot be simultaneously localized arbitrarily well in both time and frequency domains. Quantitatively, the product of the time and frequency resolutions satisfies Δt Δf ≥ 1/(4π), where equality is achieved only for a Gaussian window function. This lower bound implies that the STFT provides a fixed "area" of resolution in the time-frequency plane, preventing perfect joint localization regardless of parameter choices. The principle underscores why the STFT cannot resolve both rapid temporal changes and fine spectral details simultaneously, much like the quantum mechanical constraint on position and momentum. The shape of the window further modulates this trade-off beyond mere duration. Windows with a narrow mainlobe in the frequency domain, such as the , offer better frequency resolution by concentrating energy in a tighter spectral range but at the cost of poorer time resolution due to increased spectral leakage and less effective temporal confinement. In contrast, windows with a broader mainlobe, like the , provide a more balanced resolution, achieving the minimal uncertainty product while reducing sidelobe artifacts, though they may slightly widen the effective frequency spread for a fixed time duration. This shape-dependent impact highlights the need to select windows based on the signal's characteristics, prioritizing either temporal sharpness or spectral clarity as required. Increasing the overlap between consecutive window positions introduces redundancy in the STFT representation by generating more time-frequency data points, which can enhance visualization or interpolation but does not fundamentally improve the inherent time or frequency resolution. The core resolution limits remain governed by the window's duration and shape, as overlap merely densifies the sampling grid without altering the underlying localization constraints. This redundancy is particularly useful in synthesis methods but comes at the expense of higher computational load and storage without resolving the uncertainty bound.

Uncertainty Principle Application

The uncertainty principle in signal processing, analogous to its quantum mechanical counterpart, quantifies the inherent trade-off between time and frequency localization for a signal. For a square-integrable signal x(t) with finite energy E = \int_{-\infty}^{\infty} |x(t)|^2 \, dt, the time spread \Delta t and angular frequency spread \Delta \omega are defined as the standard deviations: \Delta t = \sqrt{ \frac{\int_{-\infty}^{\infty} (t - \mu_t)^2 |x(t)|^2 \, dt}{E} }, \quad \Delta \omega = \sqrt{ \frac{\int_{-\infty}^{\infty} (\omega - \mu_\omega)^2 |X(\omega)|^2 \, d\omega}{2\pi E} }, where \mu_t = \frac{\int t |x(t)|^2 \, dt}{E} is the mean time, \mu_\omega is the mean angular frequency, and X(\omega) is the of x(t). The principle states that these spreads satisfy (\Delta t)^2 (\Delta \omega)^2 \geq \frac{1}{4}, with equality achieved if and only if x(t) is a Gaussian function modulated by a complex exponential. In the context of the short-time Fourier transform (STFT), this principle extends to each local time-frequency representation, where the analysis window function g(t) determines the resolution. The STFT of x(t) at time t_0 and frequency \omega is X(t_0, \omega) = \int x(\tau) g^*(\tau - t_0) e^{-j \omega \tau} \, d\tau, and the uncertainty for the local spectrum around t_0 is bounded by the time-bandwidth product of the window: (\Delta t_g)^2 (\Delta \omega_g)^2 \geq 1/4, where \Delta t_g and \Delta \omega_g are the spreads of |g(t)|^2. This implies that the window cannot simultaneously provide fine time and frequency resolution, limiting the STFT's ability to localize transient events precisely. The Gaussian window achieves the minimal uncertainty bound in the STFT, saturating the inequality and providing optimal joint time-frequency concentration among all windows. For a Gaussian g(t) = (2\pi \sigma_t^2)^{-1/4} \exp(-t^2 / (4\sigma_t^2)), the product \Delta t_g \Delta \omega_g = 1/2, yielding the tightest possible localization for the local spectra. In practice, the effective resolution of the STFT satisfies \Delta t \Delta f \approx 1 / (2 \tau), where \Delta f = \Delta \omega / (2\pi) is the frequency spread in hertz and \tau is the window duration in seconds; this approximation arises from the window's second-moment spreads and highlights the fixed trade-off enforced by the choice of \tau. However, the STFT's constant window size imposes a uniform resolution across all frequencies, unlike adaptive methods such as wavelets that vary the effective window to better handle multi-scale features. In the discrete-time STFT, quantization effects further constrain the resolution, with frequency bin spacing given by $1/(N T_s) (where N is the window length in samples and T_s is the sampling interval) and time localization limited by T_s, reinforcing the uncertainty bound in finite implementations.

Spectrogram Representation

The spectrogram of a signal is defined as the squared magnitude of its short-time Fourier transform (STFT), denoted as S(\tau, \omega) = |X(\tau, \omega)|^2, where X(\tau, \omega) represents the . This quantity serves as a time-frequency energy density function, quantifying the distribution of the signal's energy across both time and frequency domains. In visualization, the spectrogram is rendered as a two-dimensional image, with the horizontal axis corresponding to time \tau, the vertical axis to frequency \omega, and the intensity or color encoding the value of S(\tau, \omega). Due to the broad dynamic range of signal amplitudes, a logarithmic scaling—such as $10 \log_{10} |S(\tau, \omega)|—is commonly applied to enhance visibility of both strong and weak components. Key properties of the spectrogram include its real-valued and non-negative nature, stemming directly from the magnitude-squared operation, which ensures S(\tau, \omega) \geq 0 everywhere. The time marginal, obtained by integrating over frequency, yields \int S(\tau, \omega) \, d\omega = \int |x(u)|^2 |w(u - \tau)|^2 \, du, representing the signal's instantaneous power density smoothed by the window's energy profile; for a unit-energy window, this approximates |x(\tau)|^2. The frequency marginal, integrating over time, approximates \int S(\tau, \omega) \, d\tau \approx |X(\omega)|^2 |W(\omega)|^2, where W(\omega) is the Fourier transform of the window function, providing an estimate of the overall power spectrum modulated by the window's frequency response. The units of the spectrogram correspond to energy (or power) per unit time per unit frequency, reflecting its role as a density in the time-frequency plane. Computationally, the spectrogram is obtained by calculating the squared magnitude of the discrete Fourier transform (DFT) for each overlapping windowed frame of the signal, typically normalized by the window's energy to ensure consistent scaling: for a frame at time index n, S(n, k) = \left| \sum_{m=0}^{N-1} x(n+m) w(m) e^{-j 2\pi k m / N} \right|^2 / \sum |w(m)|^2, where N is the frame length and k indexes frequency bins. This approach leverages efficient (FFT) implementations for practical evaluation.

Resolution Limits

Rayleigh Frequency Criterion

The Rayleigh frequency criterion establishes the minimum frequency separation required to resolve two closely spaced sinusoidal components in the short-time Fourier transform (STFT) spectrum, analogous to the resolvability limit in optical spectroscopy. Two frequencies f_1 and f_2 are considered resolvable if their separation \Delta f = |f_1 - f_2| is at least the frequency resolution of the analysis window, typically \Delta f \approx 1/T, where T is the effective duration of the window function. This limit arises because the window's Fourier transform acts as a low-pass filter, convolving with the signal's spectrum and blurring features narrower than the main lobe width. The criterion originates from Lord Rayleigh's 1879 work on optical resolution, where two point sources are resolvable if the peak of one diffraction pattern () coincides with the first minimum of the other, setting the angular resolution \theta \approx 1.22 \lambda / D for aperture diameter D and wavelength \lambda. In the STFT context, this is adapted to the half-width at half-maximum () of the main lobe in the window's frequency response, ensuring the spectral peaks of the two tones do not overlap excessively and produce a detectable dip in the combined magnitude spectrum. For a , which achieves the minimum time-frequency uncertainty, the resolution is refined to \Delta f = 0.89 / T, reflecting the full width at half maximum () of the Gaussian's frequency-domain response relative to the time-domain FWHM T. In the discrete-time STFT, frequency bins are spaced by \Delta f = f_s / N, where f_s is the sampling frequency and N is the number of samples in the window (so T = N / f_s), corresponding to a nominal bin resolution of f_s / N. However, the true resolvability, governed by the window's main lobe, can exceed single-bin separation through window design and overlap, as overlapping segments allow averaging that mitigates sidelobe interference while preserving the core resolution limit set by T. This criterion fundamentally impacts the STFT's ability to distinguish closely spaced tones; for instance, a 20 ms window (T = 0.02 s) yields \Delta f \approx 50 Hz, below which tones merge into a single broadened peak, limiting applications requiring fine spectral detail. In angular frequency terms, the Rayleigh separation is given by \Delta \omega_\text{Rayleigh} = 2\pi / T.

Gabor Limit and Examples

In his seminal 1946 work on communication theory, Dennis Gabor derived an uncertainty relation that sets a fundamental lower bound on the joint time-frequency resolution of signals. Specifically, for a signal's effective time duration \Delta t (defined as \sqrt{2\pi} times the root-mean-square deviation from the mean time) and effective frequency bandwidth \Delta f (similarly defined from the mean frequency), the product satisfies \Delta t \Delta f \geq 1/2, with equality attained for a Gaussian envelope modulated by a complex sinusoid. This bound, known as the Gabor limit, directly constrains the short-time Fourier transform (STFT), where the analysis window determines the local time and frequency spreads, preventing simultaneous arbitrary precision in both domains. The STFT of elementary signals like Gaussian-modulated sinusoids—termed —exemplifies this minimal spread. For such a signal s(t) = g(t - t_0) e^{j 2\pi f_0 (t - t_0)}, where g(t) is a Gaussian, the STFT magnitude concentrates energy in a time-frequency cell of area approximately $1/2, achieving the limit without further spreading. This optimal localization underscores why the Gaussian window is preferred in STFT for balancing resolution, as deviations from Gaussian shapes increase the product beyond $1/2. A practical example is the linear chirp signal, whose instantaneous frequency varies linearly with time, such as s(t) = e^{j \pi \mu t^2} for rate \mu. The STFT yields a spectrogram with a diagonal ridge tracing the frequency sweep, but the fixed window causes smearing: short windows (e.g., 5 ms) capture rapid changes near high frequencies yet broaden low-frequency portions due to poor resolution (\Delta f \approx 200 Hz), while longer windows (e.g., 50 ms) sharpen the ridge but blur the trajectory's curvature, producing artifacts like thickened or jagged lines in the time-frequency plane. This illustrates how the Gabor limit manifests as unavoidable distortion in representing non-stationary components. For closely spaced sinusoids, consider a two-tone signal with equal-amplitude components at 440 Hz and 450 Hz. Using a 10 ms Hamming window, the frequency resolution is roughly $1/T = 100 Hz, exceeding the 10 Hz separation, so the STFT spectrogram merges the tones into a single broadened peak rather than resolving two distinct horizontal ridges. Smearing artifacts appear as sidelobes or a filled valley between peaks, emphasizing that components closer than the window's inverse duration remain unresolvable per the limit. An impulse train s(t) = \sum_{n} \delta(t - n\tau), with period \tau, has a Fourier transform as lines at harmonics of $1/\tau. The STFT, however, localizes impulses in time but widens frequency lines to \Delta f \approx 1/T_w (where T_w is window length), or vice versa, yielding spectrograms with vertical streaks of finite width instead of delta-like lines. For \tau = 10 ms and T_w = 20 ms, time localization smears across impulses while frequency lines broaden, creating cross-term artifacts that highlight the trade-off bound by \Delta t \Delta f \geq 1/2. Extensions like adaptive windows approach the limit for specific signals without surpassing it. The chirplet transform, using chirp-modulated Gaussians, better aligns with linear frequency sweeps in chirps, reducing smearing in the spectrogram-like representation compared to fixed STFT windows, yet the minimal product remains $1/2 for optimal cases.

Applications

Audio and Speech Analysis

The short-time Fourier transform (STFT) plays a central role in audio and speech analysis by providing a time-frequency representation that reveals evolving spectral content essential for processing non-stationary signals like human voice. This transform divides the signal into overlapping frames and computes the Fourier spectrum for each, allowing analysts to track features such as harmonic structures and resonant frequencies over time. In practice, the STFT facilitates tasks from feature extraction to real-time processing in systems like automatic speech recognition and audio editing software. Pitch detection in speech signals often relies on the autocorrelation of individual STFT frames to identify periodicities corresponding to the fundamental frequency. This approach exploits the fact that voiced speech exhibits strong harmonic correlations, which the autocorrelation function enhances by suppressing aperiodic noise. A seminal method involves computing the short-time autocorrelation from STFT magnitudes or directly in the time domain per frame, enabling robust estimation even in noisy conditions. Alternatively, harmonic summation in the spectrogram—derived from STFT magnitudes—detects pitch by summing energy along expected harmonic lines, improving accuracy for polyphonic audio. The spectrogram representation from the STFT is particularly valuable for visualizing these pitch contours in speech analysis. Formant analysis utilizes peaks in the spectral envelope of STFT frames to identify resonant frequencies of the vocal tract, which are crucial for vowel recognition and speaker characterization. These formants appear as dark bands in the spectrogram and are extracted by fitting envelopes to the magnitude spectra of short-time frames, distinguishing vowel qualities based on their frequency positions. For instance, the first two formants (F1 and F2) dominate vowel spectra, and STFT-based tracking allows precise measurement across utterances, aiding in phonetic analysis and synthesis. Noise reduction in audio and speech employs spectral subtraction on STFT magnitudes, where an estimate of the noise spectrum is subtracted from the noisy signal's magnitude spectrum, while the phase is retained from the original for reconstruction. This technique assumes additive noise uncorrelated with speech, effectively attenuating broadband interference like background hum or babble without distorting the time-domain waveform. Introduced as a computationally efficient method, it has become foundational in enhancing speech intelligibility, though variants address artifacts like musical noise through over-subtraction factors. In audio compression, transforms closely related to the STFT, such as the modified discrete cosine transform (MDCT), enable perceptual coding by concentrating signal energy into fewer coefficients for efficient bitrate reduction. The MDCT, akin to the STFT in its use of overlapping windows and time-frequency decomposition, supports block-switching for transient signals and psychoacoustic masking models in standards like , achieving high-fidelity reconstruction at low bitrates by discarding inaudible components. This lapped transform's similarity to STFT ensures seamless integration in hybrid coders for music and speech. Real-time applications include voice activity detection (VAD), where STFT frame energies, particularly in low-frequency bands (below 1 kHz), distinguish speech from silence or noise based on sudden increases in spectral power. Low-frequency emphasis captures voicing energy effectively, with thresholds applied per frame to trigger detection in systems like teleconferencing, reducing computational load by processing only active segments. Energy-based VAD using STFT has proven reliable in low-noise environments, supporting features like endpointing in speech recognizers.

Radar and Sonar Signal Processing

In radar and sonar systems, the short-time Fourier transform (STFT) plays a critical role in processing non-stationary signals to detect and analyze transient echoes from distant targets. By providing a time-frequency representation, the STFT enables the extraction of instantaneous frequency components, which is essential for handling the dynamic nature of electromagnetic or acoustic waves reflected from moving objects in noisy environments. This capability has been particularly valuable in military applications for identifying radar emitters through their unique modulation patterns in the time-frequency domain. In radar systems employing linear frequency-modulated (chirp) waveforms, pulse compression is achieved through matched filtering, which correlates the received echo with a replica of the transmitted chirp to compress the wide pulse into a narrow one, achieving high range resolution, often on the order of meters, even for low-power transmissions. The STFT can analyze the time-varying frequency content of such chirp signals to support coherent integration in pulse Doppler radar, enhancing detection of weak targets amid noise. For Doppler processing, the STFT generates spectrograms that track time-frequency trajectories of moving targets, capturing Doppler shifts induced by radial velocity. In particular, micro-Doppler signatures—arising from vibrations or rotations of target components like rotor blades on unmanned aerial vehicles—are visualized as periodic sidebands around the main Doppler ridge in the STFT domain, enabling classification of target type and motion. These tracks facilitate the discrimination of maneuvering targets from clutter by isolating velocity-specific energy concentrations, with applications in airborne radar for real-time threat assessment. Clutter rejection leverages the STFT spectrogram to separate target returns from stationary interference, such as sea or ground clutter, which exhibit low or zero Doppler spread compared to moving targets. By applying Doppler filters to the time-frequency representation, stationary clutter is suppressed while preserving target energy, improving signal-to-clutter ratios in airborne or surface radar systems. For instance, in maritime surveillance, STFT-based analysis exploits the distinct spectral signatures to filter sea clutter, allowing detection of slow-moving vessels. This technique briefly references the resolution limits for target separation, where window length trades time and frequency precision to resolve closely spaced echoes. In active sonar systems, the STFT aids echo detection in reverberant underwater channels affected by multipath propagation, where direct and reflected paths create overlapping arrivals. The transform converts broadband pings and echoes into spectrograms, highlighting target-induced delays and Doppler shifts against multipath interference, which appears as dispersed low-energy replicas. Constant false alarm rate processing on the STFT output then thresholds the time-frequency plane to detect targets reliably, even in multipath-dominated environments like shallow waters. This approach supports submarine or mine detection by isolating primary echoes from channel-induced artifacts.

Biomedical Signal Processing

In biomedical signal processing, the short-time Fourier transform (STFT) is widely employed to analyze non-stationary physiological signals, providing time-localized frequency representations that reveal dynamic changes in bioelectric activity essential for medical diagnostics. By segmenting signals into overlapping windows and applying the Fourier transform to each, STFT generates spectrograms that highlight frequency modulations over time, aiding in the identification of pathological patterns without invasive procedures. This approach is particularly valuable for signals like electroencephalograms (EEG), electrocardiograms (ECG), and electromyograms (EMG), where temporal evolution of spectral content informs clinical assessments. In EEG analysis, STFT facilitates the examination of event-related potentials (ERPs) by quantifying time-frequency power changes, such as increased theta-band activity during cognitive tasks or gamma oscillations in response to stimuli, enabling precise localization of neural events. For instance, STFT spectrograms have been used to decompose ERP data into frequency bands, revealing subtle differences in auditory processing that correlate with neurological disorders like . This method enhances detection of transient brain responses, supporting applications in epilepsy diagnosis through epileptic spike identification via power spectral shifts. The STFT's fixed window limits resolution for ultra-short events, aligning with the uncertainty principle's trade-off between time and frequency precision. For ECG signals, STFT supports QRS complex detection by emphasizing high-frequency bursts in the spectrogram corresponding to ventricular depolarization, achieving high sensitivity in noisy recordings from ambulatory monitors. Additionally, STFT-derived time-frequency features quantify heart rate variability (HRV) by tracking low-frequency (0.04–0.15 Hz) and high-frequency (0.15–0.4 Hz) components, which indicate autonomic nervous system balance and predict risks like arrhythmias. Algorithms utilizing adaptive thresholds on STFT spectrograms enable QRS delineation in real-time settings. In EMG analysis, STFT identifies muscle activation onset through time-localized frequency bursts, typically in the 20–500 Hz range, where sudden increases in spectral power signal the start of motor unit firing during voluntary contractions. This is crucial for gait analysis and rehabilitation, as STFT spectrograms reveal shifts in mean frequency during fatigue, with decreases indicating metabolic changes in muscle fibers. Studies on lower limb EMG have shown STFT effectively differentiates age-related activation patterns, with older adults exhibiting broader spectral spreads due to slower motor unit recruitment. STFT contributes to sleep staging by analyzing overnight EEG spectrograms for rhythmic patterns, such as alpha (8–12 Hz) dominance in relaxed wakefulness and delta (0.5–4 Hz) prevalence in deep non-REM stages, automating classification in multi-channel recordings. By windowing 30-second epochs, STFT captures transitions between stages, aiding in the diagnosis of disorders like insomnia through quantified rhythm disruptions. Convolutional neural networks trained on STFT images of EEG have further improved staging precision by integrating spectral-temporal features. Post-2020 advances have integrated STFT with machine learning for anomaly detection in wearable biosensors, converting physiological signals into spectrogram images for input to convolutional neural networks that identify irregularities like seizures in EEG or arrhythmias in ECG. For example, hybrid STFT-ML frameworks on wearable EEG devices enable real-time seizure onset detection by flagging anomalous frequency bursts, reducing false alarms through ensemble classifiers. These developments enhance remote monitoring in telemedicine, with applications in cognitive health via unobtrusive biosensors.

Implementation Methods

Direct Computation Approach

The direct computation approach to the Short-time Fourier transform (STFT) involves sequentially processing the input signal by extracting overlapping or non-overlapping segments, applying a window function to each segment, and then calculating the discrete Fourier transform (DFT) of the windowed segment using the explicit summation formula. For a discrete-time signal x of length L, the process begins with the discrete STFT definition, where for each frame index m = 0, 1, \dots, M-1 and hop size H, the windowed segment is formed as y_m = x[n + mH] w for n = 0 to N-1, with w the window of length N. The DFT of this segment is then computed directly as X(m, k) = \sum_{n=0}^{N-1} y_m e^{-j 2\pi k n / K}, \quad k = 0, 1, \dots, K-1, where K is the number of frequency bins (often K = N). This summation is performed for every frequency bin k and every frame m, without relying on any fast approximation algorithms. The computational complexity of this method is O(M N K) operations, as each of the M frames requires O(N K) arithmetic operations for the double summation over time and frequency indices (assuming N \approx K). With M \approx L / H, this scales as O(L N^2 / H) for typical choices where the window length N equals the DFT size K, making it prohibitive for long signals or large N. This approach imposes constraints such as the need to store or buffer the entire signal for frame extraction, though streaming is possible for memory efficiency; however, it requires significant runtime for signals exceeding a few thousand samples or when N > 512. It is suitable only for short signals, small DFT sizes, or scenarios where implementation simplicity outweighs efficiency. A simple pseudocode representation in a high-level language like illustrates the process:
def direct_stft(x, w, H, N, K):
    L = len(x)
    M = (L - N) // H + 1
    X = np.zeros((M, K), dtype=[complex](/page/Complex))
    for m in range(M):
        # Extract and [window](/page/Window) segment (zero-pad if needed)
        start = m * H
        ym = np.zeros(N)
        end = min(start + N, L)
        ym[:end - start] = x[start:end] * w[:end - start]
        # Direct DFT [summation](/page/Summation)
        for k in range(K):
            for n in range(N):
                X[m, k] += ym[n] * np.exp(-1j * 2 * np.pi * k * n / K)
    return X
This loop explicitly computes each transform via nested summations, highlighting the method's straightforward but resource-intensive nature. The direct computation approach is primarily used in educational contexts to illustrate STFT fundamentals or for processing very short data segments where computational speed is not a bottleneck and code simplicity is prioritized.

FFT-Based Acceleration

The FFT-based acceleration of the short-time Fourier transform (STFT) replaces the (DFT) for each windowed signal frame with the (FFT), leveraging efficient algorithms to compute frequency content rapidly. This approach, rooted in the Cooley-Tukey radix-2 FFT, decomposes the DFT into logarithmic stages of operations, enabling practical computation of spectrograms for long signals. For an STFT with M frames of length N, the per-frame complexity drops from O(N2) in direct DFT evaluation to O(N log N) using FFT, yielding a total complexity of O(M N log N). In implementation, frame extraction handles overlap through a hop size R (typically N/2 for 50% overlap), ensuring continuity across frames while the FFT is applied independently to each windowed segment. Pre-processing often includes zero-padding the windowed frame from length MN to the next power of 2 for N, optimizing radix-2 FFT efficiency without altering the underlying but interpolating frequency bins for finer . Post-processing may involve scaling to account for padding and window normalization, preserving in the . The radix-2 FFT constrains N to powers of 2 for optimal performance; for arbitrary or prime lengths, Bluestein's reformulates the DFT as a computable via larger FFTs, retaining O(N log N) complexity per frame. This method is widely supported in software libraries, such as SciPy's signal.stft function in , which employs NumPy's FFT with configurable and overlap, and MATLAB's stft in the Toolbox, which defaults to power-of-2 FFT sizes with zero-padding. This acceleration forms the basis for real-time audio applications, where a 1024-point FFT with 512-sample hop size balances computational load and at 44.1 kHz sampling rates, enabling low-latency generation in systems like workstations.

Recursive and Sliding Algorithms

Recursive and sliding algorithms for the short-time Fourier transform (STFT) enable efficient computation by reusing results from overlapping frames, updating the transform incrementally as new samples arrive rather than recomputing the full DFT each time. These methods are particularly valuable in scenarios requiring low-latency , such as real-time signal , where the hop size is small relative to the window length. By exploiting the sliding window nature of the STFT, they achieve independent of the window size for updates, contrasting with full-frame approaches that with the FFT length. The sliding discrete Fourier transform (SDFT), also known as the sliding window DFT, is a foundational recursive technique for STFT computation. It updates each frequency by subtracting the contribution of the outgoing sample from the previous window, adding the incoming sample, and applying a rotation to account for the window shift, resulting in O(1) operations per per . This approach maintains the DFT coefficients across consecutive with minimal arithmetic—typically one complex multiplication and addition per —making it suitable for continuous monitoring. The SDFT was formalized in literature as an efficient alternative to periodic FFT evaluations for overlapping windows. For applications needing only a of bins, the provides a recursive (IIR) filter implementation per , computing individual DFT terms without evaluating the full transform. Each acts as a second-order tuned to the target , accumulating and through recursive updates that require fewer multiplications than a complete DFT for sparse spectra. This makes it ideal for tasks like pitch tracking in audio signals, where monitoring a few harmonics suffices, and it integrates into STFT frameworks by applying the filter to windowed segments. The algorithm originates from efficient trigonometric series evaluation and has been adapted for tone detection in STFT-based systems. More advanced recursive STFT implementations leverage polyphase networks or overlap-save methods akin to Blackman-Tukey spectral estimation, decomposing the into polyphase components for parallel filtering and efficient overlap handling. In polyphase structures, the STFT is realized as a where and reduce redundancy, enabling recursive updates via noble identities in multirate systems. Overlap-save techniques, drawing from acceleration, process incoming samples by saving valid portions of prior transforms, minimizing boundary effects in overlapping frames. These methods are employed in hardware-efficient spectrum analyzers, supporting embedded applications like monitoring. A key challenge in these recursive algorithms is under finite-precision arithmetic, as accumulated phase rotations in SDFT can lead to drift or without . arises from errors in recursive multiplications, potentially causing bin values to grow unbounded; solutions include factors, periodic , or exact arithmetic in fixed-point implementations. Guaranteed-stable variants incorporate correction terms to bound errors, ensuring long-term reliability in embedded systems. These algorithms find prominent use in embedded systems, such as digital spectrum analyzers for audio and RF signals, where low computational overhead supports continuous operation on resource-constrained like DSP chips. For instance, SDFT enables per-sample updates in portable devices for live frequency tracking, outperforming FFT in latency-critical scenarios.

Advanced Techniques and Comparisons

The Chirp Z-Transform (CZT) enables the computation of the STFT at arbitrary frequency points by evaluating the Z-transform along a contoured path in the complex plane through a chirp-modulated convolution, which is efficiently implemented using the FFT. This approach maintains an O((N + M) log(N + M)) time complexity, where N is the signal length and M the number of output frequency points, comparable to standard FFT-based methods but with support for non-uniform frequency sampling. Originally proposed for general Z-transform evaluation, the CZT has been adapted for short-time spectral analysis to provide flexible frequency resolution without requiring uniform grids. It is particularly advantageous for applications requiring zoomed-in spectra, such as high-resolution analysis in narrow frequency bands, where traditional FFT methods waste computational resources on irrelevant regions through zero-padding. For instance, in harmonic analysis of non-stationary signals, CZT achieves finer resolution (e.g., 0.625 Hz versus 12.5 Hz with an equivalent FFT window) while preserving efficiency. Other advanced techniques enhance STFT flexibility and performance. The weighted overlap-add () method applies an additional synthesis window after the inverse STFT to accommodate variable hop sizes, ensuring perfect signal reconstruction even when overlaps do not satisfy constant overlap-add () conditions exactly. This is useful for adaptive processing where hop lengths vary to balance time resolution and computational load. Additionally, GPU parallelization leverages the inherent parallelism of STFT operations—such as batched FFTs across time frames—for applications, achieving speedups of up to 35x over CPU implementations in large-scale audio processing pipelines.
MethodTime Complexity (per window)Memory UsageAccuracy/Notes
Direct ComputationO()O(N)Exact; simple but inefficient for large N.
FFT-BasedO(N log N)O(N)Exact; balanced for uniform grids; widely used baseline.
Sliding DFT (SDFT)O(N) (for hop size 1)O(N)Near-exact with stable implementations; ideal for real-time, low-power streaming with minimal updates.
Chirp Z-Transform (CZT)O((N + M) log(N + M))O(N + M); excels in non-uniform/zoomed frequencies; flexible but slightly higher constant factors than FFT.
In the 2020s, modern trends emphasize integration with frameworks through vectorized libraries like PyTorch's torch.stft, which supports batched, GPU-accelerated computations for seamless embedding in pipelines, such as audio classification or source separation models.

References

  1. [1]
    The Short-Time Fourier Transform - Stanford CCRMA
    The Short-Time Fourier Transform (STFT) (or short-term Fourier transform) is a powerful general-purpose tool for audio signal processing.
  2. [2]
    [PDF] Short-Time Fourier Analysis
    In an effort to correct this deficiency, Dennis Gabor (1946) adapted the Fourier transform to analyze only a small section of the signal at a time -- a ...
  3. [3]
    Local time-frequency analysis and short time Fourier transform
    The resulting local time-frequency analysis procedure is referred to as (continuous) short time Fourier transform or windowed Fourier transform.
  4. [4]
    Mathematical Definition of the STFT
    ### Mathematical Definition of STFT
  5. [5]
    [PDF] B3. Short Time Fourier Transform (STFT) - Faculty
    In what follows we introduce the Short Time Fourier Transform (STFT) and its applications to the analysis of signals. 2. The Short Time Fourier Transform.
  6. [6]
    What is the Gabor uncertainty principle? - Agile Scientific
    Jan 15, 2014 · Dennis Gabor (1946), inventor of the hologram, was the first to realize that the uncertainty principle applies to signals. Thanks to wave- ...Missing: mechanics | Show results with:mechanics<|control11|><|separator|>
  7. [7]
    Discrete Short-Time Fourier Transform (STFT)
    To recover the hidden time information, Dennis Gabor introduced in the year 1946 the short-time Fourier transform (STFT). Instead of considering the entire ...
  8. [8]
    None
    ### Summary of Continuous Short-Time Fourier Transform (STFT) from the Document
  9. [9]
    Time Fourier Transform - an overview | ScienceDirect Topics
    In the year 1946, a Hungarian electrical engineer and physicist, Dennis Gabor (1900–1979), fostered the idea in developing Short Time Fourier Transform (STFT) ...
  10. [10]
    [PDF] THEORY OF COMMUNICATION* By D. GABOR, Dr. Ing., Associate ...
    Because of its self-reciprocal character, the probability signal has always played an important part in the theory of Fourier transforms. In three recent papers ...
  11. [11]
    On the use of windows for harmonic analysis with the discrete ...
    On the use of windows for harmonic analysis with the discrete Fourier transform. Abstract: This paper makes available a concise review of data windows and ...
  12. [12]
    Kaiser Window | Spectral Audio Signal Processing - DSPRelated.com
    The Kaiser window is a simple approximation to the DPSS window based on Bessel functions. Its beta parameter controls side-lobe level and main-lobe width.Missing: properties | Show results with:properties
  13. [13]
    The Short-Time Fourier Transform | Spectral Audio Signal Processing
    The Short-Time Fourier Transform (STFT) is a powerful tool for audio signal processing, defining time-frequency distributions for any signal.
  14. [14]
    [PDF] 3. Short-Time Fourier Transforms
    This implies that in principle we can obtain the time function by multiplying the vari- ous short-time Fourier transform coefficients X[n,k] by the function ...
  15. [15]
    [PDF] EE 438 Digital Signal Processing with Applications: Short Time ...
    Mar 31, 2000 · We define the short-time continuous-time Fourier transform (STCTFT) ... simply have the convolution of the Fourier transforms of these two terms.
  16. [16]
    [PDF] Selected Advanced Topics in Digital Signal Processing
    Aug 27, 2020 · Generalizing, we can define the continuous-time analog of the STFT as. Xc(t,Ωk) = Z ∞. −∞ xc(τ)wc(t −τ)e−jΩkτdτ. (3.39) where wc(t) is a ...Missing: textbook | Show results with:textbook
  17. [17]
    [PDF] Short Time Fourier Transform, Integral Wavelet Transform
    We will discuss Short Time Fourier transforms (STFT), integral wavelet transforms ... By using the Parseval Identity, the recurrence relation Nn+1 = Nm ∗ N1, and.
  18. [18]
    [PDF] A Unified Approach to Short-Time Fourier Analysis and Synthesis
    Nov 11, 1977 · It is the purpose of this paper to unify the various methods of analysis and synthesis, and to show the effects of modifying the short-time ...
  19. [19]
    [PDF] L6: Short-time Fourier analysis and synthesis
    • Discrete-time Short-time Fourier transform. – The Fourier transform of the windowed speech waveform is defined as. X n,𝜔 = x m w n − m e−j𝜔n.
  20. [20]
    [PDF] NOTES ON SHORT-TIME FOURIER TRANSFORMS I. Introduction
    This observation motivated the development of the short-time Fourier transform (STFT). ... windows are used, on the other hand, a 50% overlap will be needed, ...
  21. [21]
    [PDF] ESE 5310: Digital Signal Processing
    window type, DFT length (zero-padding). ▫ Get accurate representation of DFT. ❑ Time-dependent Fourier transform. ▫ Aka short-time Fourier transform.
  22. [22]
    Zero Padding Applications - Stanford CCRMA
    Zero padding in the time domain is used extensively in practice to compute heavily interpolated spectra by taking the DFT of the zero-padded signal. Such ...Missing: short- | Show results with:short-
  23. [23]
    Overlap-Add (OLA) STFT Processing - DSPRelated.com
    This chapter discusses use of the Short-Time Fourier Transform (STFT) to implement linear filtering in the frequency domain.
  24. [24]
    [PDF] MUS421 Lecture 8A FFT Signal Processing: The Overlap-Add (OLA ...
    Jun 27, 2020 · In the weighted overlap add (WOLA) method, we apply a second window after the inverse DFT. Such a window can be called a “synthesis window”, “ ...
  25. [25]
    [PDF] Time-Frequency Analysis of Biophysical Time series
    time/position and in frequency/momentum. There exists a lower bound to the Heisenberg product: Δt Δf ≥ 1/(4π) or Δf ≥ 1/(4πΔt). Δf = 1Hz, Δt = 80 msec or.
  26. [26]
    [PDF] Determining Shock and Detonation Wave Propagation Time Using ...
    Jun 25, 2009 · The uncertainty principle is a modification of the Heisenberg uncertainty principle applied for signal processing. The uncertainty principle ...
  27. [27]
    Uncertainty principles for the Fourier and the short ... - AIP Publishing
    Nov 1, 2021 · The aim of this paper is to establish a few uncertainty principles for the Fourier and the short-time Fourier transforms.INTRODUCTION AND... · II. THE SHORT-TIME... · III. PROOFS OF MAIN RESULTS
  28. [28]
    Systematic analysis of uncertainty principles of the local polynomial ...
    Dec 16, 2014 · It was shown that when the Gaussian window is used to segment the signal, the uncertainty product of the LPFT is time-independent if the ...
  29. [29]
    [PDF] Lecture 16 Limitations of the Fourier Transform: STFT
    Understand the logic behind the Short-Time Fourier Transform (STFT) in order to ... For a function f(t), the STFT is defined as follows: ˆ f(t, u) = ∫. 1. 1.
  30. [30]
    [PDF] Introduction to Time-Frequency Distributions
    Jan 19, 2010 · One can define an energy density, called the spectrogram, by: PS(t,ω) = |S(t,ω)|2 = |. Z f(τ)g(τ − t)e−jωτ dτ|2. (3). The spectrogram ...
  31. [31]
    [PDF] Time Frequency Signal Analysis and Processing - Semantic Scholar
    Chapter 1: Time-Frequency Concepts (B. Boashash). 3. Overview. 3. 1.1 The Need for a Time-Frequency Distribution (TFD).
  32. [32]
    XXXI. Investigations in optics, with special reference to the ...
    Investigations in optics, with special reference to the spectroscope. Lord Rayleigh FRS. Pages 261-274 | Published online: 13 May 2009.
  33. [33]
    [PDF] Superlets: time-frequency super-resolution using wavelet sets
    Sep 8, 2019 · 1A, left). For a given window size, the STFT has fixed frequency resolution ... maximum (FWHM) of the frequency and temporal responses measured at ...<|control11|><|separator|>
  34. [34]
    [PDF] Sample manuscript showing specifications and style
    simplest time-frequency representation is the Short-time Fourier Transform (STFT). ... to Rayleigh criterion. In other words, the ... a) Short-time Fourier ...
  35. [35]
    [2506.20864] The Rayleigh Criterion: Resolution Limits of ... - arXiv
    Jun 25, 2025 · The Rayleigh criterion determines the resolution limit of a periodogram, which is the minimum frequency separation required to barely resolve two sinusoids.
  36. [36]
  37. [37]
    [PDF] Gabor Expansion and Frame Theory
    Hence, the STFT suffers from a fundamental resolution tradeoff and resolution limitation. Once an analysis window is chosen, the time and frequency resolutions ...
  38. [38]
    [PDF] PARAMETER ESTIMATION OF CHIRP SIGNAL USING STFT - IJAET
    This paper suggested a technique called short time Fourier transform (STFT) for the parameter estimation of chirp signal in intercept SONAR.
  39. [39]
    [PDF] time-frequency analysis of musical signals using the phase
    Sep 6, 2013 · Here, we used a noisy synthetic sound with three sinusoids with frequency 440 Hz, 445 Hz and 450 Hz. In these figures, the amplitude ...
  40. [40]
    Applications of the STFT | Spectral Audio Signal Processing
    Applications of the STFT. This chapter briefly tours selected applications involving spectral audio signal processing, with associated examples in matlab.
  41. [41]
  42. [42]
  43. [43]
    Coherent‐like integration for PD radar target detection ... - IET Journals
    Jan 1, 2020 · The PD radar returned signals are first performed by stretch processing. Stretched signals are analysed by short-time Fourier transform (STFT) ...
  44. [44]
    Radar Detection-Inspired Signal Retrieval from the Short-Time ...
    This paper presents a novel adaptive algorithm for multicomponent signal decomposition from the time–frequency (TF) plane using the short-time Fourier ...
  45. [45]
  46. [46]
  47. [47]
    SEA CLUTTER SUPPRESSION BASED ON SHORT-TIME ...
    SEA CLUTTER SUPPRESSION BASED ON SHORT-TIME FOURIER TRANSFORM IN AN AIRBORNE RADAR SYSTEM ... A Novel Sea Clutter Rejection Algorithm for Spaceborne 1Multi- ...
  48. [48]
  49. [49]
  50. [50]
    Review of biomedical signal and image processing - PubMed Central
    Sep 8, 2013 · The authors lead into the concept by describing the short-time Fourier transform, which is helpful and clear. They show how windowing at short ...
  51. [51]
    Time-frequency analysis of event-related potentials associated with ...
    Mar 1, 2018 · ... related potential data into different frequency bands using short-time Fourier transform. The results identified an insignificant difference ...
  52. [52]
    The applied principles of EEG analysis methods in neuroscience ...
    Dec 19, 2023 · Epileptic seizure classification of EEG time-series using rational discrete short-time Fourier transform. IEEE Trans Bio-Med Eng. 2015;62(2): ...
  53. [53]
    Automatic QRS Detection and Segmentation Using Short Time ...
    In this paper, we present an automated algorithm for detecting QRS waves and segmenting ECG signal into separate beats using short time Fourier transform (STFT) ...
  54. [54]
    A Method for QRS Delineation Based on STFT Using Adaptive ...
    In this work, it is proposed to utilize an adaptive threshold technique on spectrogram computed using Short Time Fourier Transform (STFT) for QRS complex ...
  55. [55]
    ECG signal feature extraction trends in methods and applications
    Mar 8, 2023 · Short-Time fourier transform (STFT). The STFT can be used to compute and analyze the energy distribution of the ECG signal. It is essentially used to compute ...
  56. [56]
    Techniques of EMG signal analysis: detection, processing ...
    The common type of time frequency distribution is the Short-time Fourier Transform (STFT). According to Davies and Reisman (28) (1994), the major difficulty ...
  57. [57]
    Analysis of aging effect on lower limb muscle activity using short ...
    May 8, 2023 · Analysis of aging effect on lower limb muscle activity using short time Fourier transform and wavelet decomposition of electromyography signal ...
  58. [58]
    Automated Sleep Stages Classification Using Convolutional Neural ...
    Feb 10, 2023 · Third panel from the left: short-time Fourier transform of the raw EEG, where the x-axis is associated with the time (seconds), and the y ...
  59. [59]
    Automatic Sleep Staging Method Using EEG Based on STFT and ...
    Dec 30, 2024 · Initially, high-dimensional time-frequency features of EEG signals are extracted using a short-time Fourier transform ... sleep stages. The ...
  60. [60]
    SeizFt: Interpretable Machine Learning for Seizure Detection Using ...
    Aug 2, 2023 · This work presents SeizFt—a novel seizure detection framework that utilizes machine learning to automatically detect seizures using wearable SensorDot EEG data ...
  61. [61]
    Ensemble learning for biomedical signal classification - Nature
    Jul 1, 2025 · The Short-Time Fourier Transform (STFT) is a popular technique for converting biological signals into spectrograms, providing a time-frequency representation ...
  62. [62]
    Role of deep learning in cognitive healthcare: Wearable signal ...
    Apr 22, 2025 · Deep Learning (DL) offers promising solutions for analyzing wearable signals and gaining valuable insights into cognitive disorders.
  63. [63]
    [PDF] 1. Calculation of the DFT 2. The Fast Fourier Transform algorithm 3 ...
    The Fast Fourier Transform. 1. Calculation of the DFT. 2. The Fast Fourier Transform algorithm. 3. Short-Time Fourier Transform. Page 2. 2013-11-27. Dan Ellis.<|control11|><|separator|>
  64. [64]
    [PDF] Time-Frequency Analysis Math 211A—Spring 2006 - UCI Mathematics
    The direct computation using (6) involves. (N − 1)2 multiplications and N2 ... 6S is for “short-time Fourier transform. 17. Page 21. TF: the (time ...
  65. [65]
    [PDF] Lecture 5: Short-Time Fourier Transform and Filterbanks
    The short-time Fourier Transform (STFT) is the Fourier transform of a short part of the signal. We write either X(ωk,m) of X[k,m] to mean:.
  66. [66]
    Zero-Padding to the Next Higher Power of 2
    ... FFT with any window length $ M$ . When $ M$ is not a power of $ 2$ , we append enough zeros to make the FFT size $ N>M$ be a power of $ 2$ . In Matlab and ...
  67. [67]
    stft — SciPy v1.16.2 Manual
    STFTs can be used as a way of quantifying the change of a nonstationary signal's frequency and phase content over time.Stft · Previous scipy.signal.stft · ShortTimeFFT · 1.11.4
  68. [68]
    stft - Short-time Fourier transform - MATLAB - MathWorks
    s = stft(x) returns the Short-Time Fourier Transform (STFT) of x. example s = stft(x,fs) returns the STFT of x using sample rate fs.Istft · Dlstft · Stft
  69. [69]
    FFT Processing in JUCE - audiodev.blog
    Sep 5, 2023 · A common question that comes up is how to do FFT processing on real-time audio. ... 1024 is a good default size but 512 and 2048 are also common.
  70. [70]
    Understanding and Implementing the Sliding DFT - Eric Jacobsen
    Apr 23, 2015 · The Sliding DFT algorithm provides frequency domain updates on a per-sample basis with significantly fewer computations than the FFT for each update.
  71. [71]
    A short-time spectrum analyzer with polyphase-network and DFT
    Abstract. Two algorithms for digital spectral analysis are compared: the conventional DFT, and a generalized polyphase principle, which is implemented by ...
  72. [72]
    [PDF] Polyphase-Discrete Fourier Transform Spectrum Analysis for the ...
    Nov 15, 1991 · It is the intent of this article to propagate the theory of polyphase-DFT fil- ter banks as a technique related to tile use of windows with ...Missing: Short- | Show results with:Short-
  73. [73]
  74. [74]
    [PDF] A Guaranteed Stable Sliding Discrete Fourier Transform Algorithm to ...
    The Sliding DFT performs an N point Discrete Fourier. Transform on Time Samples within a Sliding Window ... The Sliding DFT reduce the processing time as compared ...