Signals and Systems


Course notes on continuous and discrete signals, LTI systems, Fourier/STFT, sampling, filters, Laplace and z transforms, and the bridge to speech features.
The objective in Signals and Systems is not formula memorisation, but a coherent view of the relationship between time and frequency and of how a system transforms a signal.
Unit 1: Signal and System Foundations
1.1. Signals and Systems
A signal carries information as a function of one or more independent variables; continuous-time signals are written x(t) and discrete-time signals x[n]. A system maps an input to an output:
x -> system -> yA system may be memoryless, causal, BIBO stable, time invariant, linear, or invertible. Linearity requires both additivity and homogeneity; y=2x+3 is not linear, while y[n]=x[n-1] is causal but has memory.
1.1.1. Elementary Signals
In discrete time:
δ[n] = u[n] - u[n-1]
u[n] = Σ(k=-∞..n) δ[k]In continuous time the Dirac impulse is defined by its action under integration:
u(t) = ∫(-∞..t)δ(τ)dτ
δ(t) = du(t)/dt
∫x(t)δ(t-t0)dt = x(t0)The complex exponential,
e^(jωt) = cos(ωt) + j sin(ωt)is the natural basis for sinusoidal analysis. Continuous-time e^(jω0t) is periodic for every nonzero ω0; discrete-time e^(jω0n) is periodic only when ω0/(2π) is rational, and discrete frequency itself is 2π periodic.
Energy:
E = ∫|x(t)|²dt
E = Σ|x[n]|²classifies finite-energy signals; periodic signals instead have infinite energy but finite nonzero average power.
1.1.2. Time Transformations
x(t-t0) shift
x(-t) reversal
x(at) scalingContinuous-time scaling is arbitrary; discrete-time indices must remain integral, so rate change is treated as sample removal or insertion.
Even and odd parts are:
x_e(t)=[x(t)+x(-t)]/2
x_o(t)=[x(t)-x(-t)]/21.2. LTI Systems and Convolution
A linear time-invariant system is completely determined by its impulse response.
Discrete time:
x[n]=Σx[k]δ[n-k]
y[n]=Σx[k]h[n-k]=x[n]*h[n]Continuous time:
y(t)=∫x(τ)h(t-τ)dτ=x(t)*h(t)Convolution is commutative, associative, and distributive; cascaded LTI systems convolve impulse responses and parallel systems add them.
System properties follow from h:
memoryless h(t)=Kδ(t)
causal h(t)=0, t<0
stable ∫|h(t)|dt < ∞with Σ|h[n]|<∞ in discrete time.
Step and impulse responses satisfy:
s(t)=u(t)*h(t)
h(t)=ds(t)/dt
h[n]=s[n]-s[n-1]The geometry of the integral consists of three operations: reverse the impulse response in time (h(-τ)), shift it to the output instant (h(t-τ)), and integrate its pointwise product with the input over τ. The output at any fixed t is the integral of the product of the two functions. With positive rectangular or triangular pulses, it resembles a weighted overlap area; signals with negative values require signed contributions rather than a simple unsigned area.
This geometric interpretation explains why an LTI system convolves its input with its impulse response. Matched filtering also involves reversal and shifting, but its relation to correlation must be handled explicitly; convolution and correlation should not be treated as identical operations.
1.3. Differential and Difference Equations
A continuous LTI system can be written:
Σa_k d^k y(t)/dt^k = Σb_k d^k x(t)/dt^kand a discrete system:
Σa_k y[n-k] = Σb_k x[n-k]Initial conditions determine the natural response. Recursive discrete systems generally produce IIR behavior; finite nonrecursive sums produce FIR behavior.
1.4. Eigenfunctions and Frequency Response
Complex exponentials are LTI eigenfunctions:
e^(st) -> H(s)e^(st)
z^n -> H(z)z^nThe system preserves their form and changes only a scalar coefficient, so H(jω) directly describes sinusoidal amplitude and phase change.
Unit 2: Fourier Analysis and the Frequency Domain
Frequency analysis starts by fixing observation duration, sampling, and windowing. Interpret a DFT peak in light of resolution, spectral leakage, and window effects before treating it as a physical tone. Amplitude estimates require explicit normalization and one-sided or two-sided spectrum conventions. Nonstationary signals may require local time-frequency analysis rather than a single spectrum over the entire record.
2.1. Fourier Series
For period T0 and ω0=2π/T0:
x(t)=Σa_k e^(jkω0t)
a_k=(1/T0)∫(T0)x(t)e^(-jkω0t)dta0 is the average value and real signals have conjugate-symmetric coefficients. Under the Dirichlet conditions, the series converges to the signal at continuity points and to the average of one-sided limits at jumps.
Truncation leaves an overshoot of roughly nine percent near a discontinuity; increasing the number of terms narrows but does not remove it. This is the Gibbs phenomenon.
Parseval:
(1/T0)∫|x(t)|²dt = Σ|a_k|²connects time-domain average power to harmonic power. A periodic discrete-time sequence has only N distinct harmonics.
2.2. Fourier Transform
For an aperiodic signal:
X(jω)=∫x(t)e^(-jωt)dt
x(t)=(1/2π)∫X(jω)e^(jωt)dωCore properties are:
x(t-t0) <-> e^(-jωt0)X(jω)
dx/dt <-> jωX(jω)
x(at) <-> (1/|a|)X(jω/a)
x*h <-> XH
xh <-> (1/2π)X*HCompression in time expands frequency. Perfect time limitation and perfect band limitation cannot hold simultaneously.
Useful pairs include:
δ(t) <-> 1
1 <-> 2πδ(ω)
e^(-at)u(t) <-> 1/(a+jω)
rectangle <-> sincConvolution becoming multiplication is the central simplification of LTI analysis.
2.3. DTFT, DFT, and FFT
The DTFT:
X(e^(jω))=Σx[n]e^(-jωn)
x[n]=(1/2π)∫(2π)X(e^(jω))e^(jωn)dωis 2π periodic.
For N finite samples, the DFT is:
X[k]=Σ(n=0..N-1)x[n]e^(-j2πkn/N)
x[n]=(1/N)Σ(k=0..N-1)X[k]e^(j2πkn/N)The DFT treats the block as periodic; endpoint mismatch produces spectral leakage.
The FFT is not another transform but a family of efficient DFT algorithms. Direct DFT costs about N² operations, while FFT methods require about N log N.
2.3.1. Linear and Circular Convolution
Multiplying DFT coefficients and applying the inverse DFT computes circular convolution over the transform length. Indices wrap around that length, so the tail of a finite linear convolution can be added back onto its beginning. This time-domain folding is distinct from frequency aliasing at an ADC.
For sequences of lengths N and M, with zero values outside their stated ranges, the full linear convolution occupies N+M-1 samples. To recover it through FFTs, pad both sequences to a common length K >= N+M-1:
x_pad = [x[0], ..., x[N-1], 0, ..., 0]
h_pad = [h[0], ..., h[M-1], 0, ..., 0]
y_pad = IFFT_K(FFT_K(x_pad) · FFT_K(h_pad))
y = y_pad[0 .. N+M-2]The multiplication is pointwise between complex coefficients, and the forward and inverse normalizations must agree. Padding here reserves space for the convolution tail; it does not improve the underlying spectral resolution (Smith, chapter 18).
2.3.2. Windowing
Finite observation multiplies a signal by a window. A rectangular window gives a narrow main lobe but high sidelobes; Hann and Hamming reduce leakage, while Blackman suppresses sidelobes further at the cost of a wider main lobe.
Zero padding samples the displayed spectrum more densely but does not create true frequency resolution; observation duration remains the main limit.
2.3.3. STFT
A single Fourier transform gives frequency but not time. The short-time Fourier transform analyzes overlapping windows:
STFT -> spectrogramShort windows improve time resolution and long windows improve frequency resolution. Speech, music, vibration, radar, and sonar analysis all expose this tradeoff.
When a sinusoidal component does not complete an integer number of cycles within the DFT window, its energy spreads across neighboring frequency bins. This leakage does not imply additional physical frequencies. A window can suppress sidelobes but widens the main lobe, making nearby tones harder to separate. Zero-padding creates a denser sampled frequency grid; it does not increase observation time and therefore does not improve physical frequency resolution. Window type, sampling rate, duration, and amplitude normalization must accompany interpretation of an FFT plot.
2.4. Correlation and Power Spectrum
Cross-correlation,
r_xy[k]=Σx[n]y*[n-k]measures similarity versus delay; autocorrelation compares a signal with shifted copies of itself. Delay estimation, periodicity detection, matched filtering, and synchronization use the same operation.
The Wiener-Khinchin relation connects autocorrelation and power spectral density. A finite-record periodogram has high variance; Welch's method averages windowed overlapping segments to reduce variance at the cost of frequency resolution.
2.5. From the Short-Time Fourier Transform to Speech Features
A Fourier transform describes the overall frequency content of a record, but speech and many real signals change over time. If a signal is treated as approximately stationary over a short interval, it can be divided into overlapping frames and analyzed separately. The short-time Fourier transform (STFT) builds this time-frequency representation.
samples
→ framing
→ windowing
→ STFT
→ time-frequency coefficients
→ spectrogramFrame length and hop size are chosen together. Short frames track rapid change more closely but provide poorer frequency discrimination; longer frames improve frequency discrimination while reducing time localization. Windowing helps control spectral leakage created by finite frame boundaries.
In speech processing, the STFT is often the beginning rather than the final feature representation. A common chain is:
STFT
→ power spectrum
→ Mel filter bank
→ log energy
→ log-Mel representationThe Mel scale samples frequency with a resolution closer to human auditory perception, while the logarithm compresses a wide energy range. A further transform can produce cepstral coefficients. The full path is discussed in Speech Feature Extraction with MFCC and From Spectrum to Cepstrum in Audio Processing.
Feature representations must also be kept distinct from quality metrics. SNR concerns signal and noise energy; PESQ and STOI concern perceived speech quality or intelligibility. WER and CER measure errors in the text produced by a speech recognizer. Good signal quality does not guarantee low WER, and WER alone does not describe the perceptual quality of enhanced audio.
Similarly, voice activity detection (VAD) is primarily a segmentation problem: it estimates where speech is active. MFCC or log-Mel features represent selected intervals numerically. Modern recognizers may combine or internalize these stages in different ways. Whisper Architecture in Speech Recognition Systems provides a separate end-to-end example.
The engineering lesson is that every transformation between waveform samples and model input preserves, emphasizes, or discards information. Feature extraction should therefore be understood as controlled representation of signal information, not merely as tensor preparation.
2.5.1. From Spectrum to Cepstrum, MFCC, and ASR Model Input
The time-frequency representation produced by the STFT is only the first transformation in a speech-processing chain. Taking the logarithm of spectral power converts multiplicative spectral structure into an additive form; a second transform of the log spectrum produces a cepstral coordinate system. In the real cepstrum that second transform is commonly an inverse Fourier transform. In an MFCC pipeline, spectral energy is first accumulated by a Mel filter bank, mapped to the log domain, and then usually transformed with a DCT. MFCC and the directly computed real cepstrum are therefore not the same object; their shared idea is to represent log-spectral structure with coefficients that are easier to separate and compress (Davis and Mermelstein, 1980; Oppenheim and Schafer, 2004).
This distinction matters when interpreting model input. Classical recognizers widely used hand-designed features such as MFCC, whereas some modern systems feed a log-Mel spectrogram directly to a neural network. Whisper is a clear example (Radford et al., 2022). Saying that a model does not use an explicit cepstrum therefore does not mean that Fourier analysis or log-spectral representation has become irrelevant; part of the transformation chain has simply moved before or inside the learned model.
The application path is developed at different layers in From Spectrum to Cepstrum in Audio Processing, Speech Feature Extraction with MFCC, and Whisper Architecture in Speech Recognition Systems. The co-authored Image and Audio Processing — Book Chapter brings audio feature vectors, speech recognition, and speaker recognition together in one published work. That link provides author provenance for the application context; the general claims about Fourier analysis, cepstrum, and MFCC remain grounded in independent academic sources.
2.5.2. Content-Based Audio Matching with Spectral Peaks
Local maxima in an STFT magnitude spectrogram can be selected to create a sparse time-frequency map. Since the frequency of one isolated peak is insufficiently distinctive, two nearby peaks are represented by (fₐ, fᵦ, Δt), where Δt = tᵦ − tₐ is their relative time difference. A lookup key produced from quantized frequencies and time differences can be searched in a reference index independently of the recording's absolute start time.
If the same key occurs at time tₛ in the query and tᵣ in a reference, the offset is δ = tᵣ − tₛ. Numerous distinct matches for one recording concentrated around the same δ indicate temporal alignment between excerpts. Some peaks may be lost to noise or compression while enough shared relationships remain. Wang's spectral-peak audio identification method targets partial recaptures of the same reference recording rather than speaker identity or similarity between melodies. Peak selection, fingerprint indexing, and matching performance are covered in Audio Processing and Forensic Audio Engineering.
2.6. Autocorrelation and Power Spectral Density
Autocorrelation measures similarity between a signal and delayed versions of itself as a function of lag. For a wide-sense stationary process it can be written conceptually as
R_xx(tau) = E[x(t) x(t+tau)].Under suitable conditions, the Wiener-Khinchin relation connects power spectral density to the Fourier transform of the autocorrelation. Time-domain lag structure and frequency-domain power distribution are therefore two views of the same second-order information.
A periodogram computed from a finite record is an estimate rather than the exact PSD. Window choice, segment length, and averaging create a resolution-versus-variance trade-off.
2.7. Windowed-Spectrum Amplitude Calibration and Leakage
Computing a DFT of a finite observation implicitly applies a rectangular time window. If a sinusoid does not fall exactly at a DFT-bin center, its energy spreads across nearby bins. This spectral leakage does not mean the physical source generated additional tones. Hann and Blackman windows reduce sidelobe levels at the expense of a wider main lobe. Resolving nearby frequencies and exposing weak components beneath the sidelobes of stronger components are therefore competing objectives.
Window gain matters when estimating amplitude. Let (w[n]) be a window of (N) samples and (X_w[k]) its windowed DFT. For a real-valued sinusoid exactly centered on a non-DC, non-Nyquist bin, its peak amplitude can be estimated as (A=2|X_w[k]|/sum_{n=0}^{N-1}w[n]). This amplitude correction is not the same as wideband power spectral density normalization, which involves the sampling frequency and (sum w[n]^2). Replacing a window without revisiting scaling may make otherwise identical measurements appear different.
Zero-padding increases the sampling density of the displayed DFT curve but not the inherent resolving power provided by the original observation duration. A controlled experiment should fix sampling rate and duration, then compare amplitude error, main-lobe width, and leakage under several windows. Absolute spectrogram magnitude and normalization relative to each frame's maximum are also different display contracts and should be labeled accordingly.
Unit 3: Sampling, Multirate Processing, and Filters
Choose the sampling rate from the highest signal band that must be preserved. Design the analog filter ahead of the sampler to limit aliasing from out-of-band content; a digital filter cannot reverse aliasing that has already occurred. Low-pass filtering before downsampling limits the information lost by removing samples. Verify the resulting passband and group delay against application requirements.
3.1. Sampling and Quantization
Impulse-train sampling:
x_p(t)=x(t)Σδ(t-nT)replicates the spectrum at ωs=2π/T intervals:
X_p(jω)=(1/T)ΣX(j(ω-kωs))For bandlimit ωM, the condition ωs>2ωM keeps replicas separate. 2ωM is the Nyquist rate, not the selected physical sample rate.
Aliasing makes high-frequency components appear at lower frequencies and destroys information at sampling time. A practical ADC chain is therefore:
analog -> anti-alias filter -> sample/hold -> ADC -> digital dataSampling discretizes time; quantization discretizes amplitude. An ideal B-bit ADC has 2^B levels and roughly:
SNR ≈ 6.02B + 1.76 dBfor a full-scale sinusoid. Thermal noise, clock jitter, and nonlinearity reduce practical ENOB.
Ideal reconstruction uses sinc interpolation; practical DACs use a zero-order hold followed by an analog reconstruction filter.
Sampling frequency f_s is the number of samples per second, measured in hertz. The sampling period is T_s = 1/f_s, and the time coordinate of sample n is t_n = n/f_s. At f_s = 16,000 Hz, a 20 ms frame contains 320 samples. For an N-point DFT, adjacent frequency-bin centers are separated by Δf = f_s/N; this spacing alone does not guarantee that two nearby tones are resolvable after windowing.
Band-limiting and anti-alias filtering act on the analog or higher-rate signal before the lower-rate sampling step. Interpolation that increases the number of samples cannot reconstruct frequency components already irreversibly aliased in the source. Quantization bit depth determines amplitude-axis resolution; raising the sampling frequency and increasing bit depth are distinct engineering changes.
3.2. Multirate Processing
Downsampling requires:
low-pass -> keep every Mth sampleand interpolation requires:
insert L-1 zeros -> low-passWithout prefiltering, decimation aliases the discrete spectrum. Rational L/M conversion combines both operations; polyphase structures avoid computations that would later be discarded.
3.3. Filters
The main responses are low-pass, high-pass, band-pass, band-stop, and all-pass. A brick-wall response is noncausal, so real filters trade transition width, ripple, attenuation, phase, order, and computation.
FIR:
y[n]=Σb_k x[n-k]is finite, nonrecursive, always BIBO stable, and can have exact linear phase.
IIR:
y[n]=-Σa_k y[n-k]+Σb_k x[n-k]can reach a given selectivity with fewer coefficients but requires pole-stability and numerical-sensitivity analysis.
Classical families include Butterworth, Chebyshev I and II, elliptic, and Bessel filters, each choosing a different balance among flatness, transition sharpness, and phase behavior.
3.3.1. FIR State at Startup and Block Boundaries
Let the coefficients be b[0..M-1] and all preceding input samples be zero. With b = [1/4, 1/2, 1/4] and x = [4, 0, 8, 4], direct filtering gives:
y[n] = x[n]/4 + x[n-1]/2 + x[n-2]/4
y[0] = 4/4 = 1
y[1] = 0/4 + 4/2 = 2
y[2] = 8/4 + 0/2 + 4/4 = 3
y[3] = 4/4 + 8/2 + 0/4 = 5
y[4] = 0/4 + 4/2 + 8/4 = 4
y[5] = 0/4 + 0/2 + 4/4 = 1Both arrays start at index n=0; the finite input is extended by zeros. Its four samples therefore produce the six-sample full result [1, 2, 3, 5, 4, 1]. The final two samples form the filter tail. A streaming call returning one output per input has not yet emitted that tail; processing M-1 zeros after the actual end of the record completes it.
Now split the input into [4, 0] and [8, 4]. Direct FIR processing must carry the last M-1 input samples into the next block. Resetting that history at every boundary makes the second block start with 2 instead of 3. The coefficients have not changed, but the initial conditions have. Arm CMSIS-DSP illustrates this distinction by keeping coefficients and evolving filter state in separate buffers.
Output selection also affects alignment. SciPy's convolve returns the entire result with full; same selects a centered portion with the first input's length. That portion is generally not the first N outputs of a causal stream. Compare implementations using the same initial conditions and the same retained index range.
3.4. Phase and Group Delay
Distortionless transmission requires:
H(jω)=Ke^(-jωt0)so magnitude is constant and phase is linear.
Group delay:
τg(ω)=-d∠H(jω)/dωdescribes narrowband envelope delay. Frequency-dependent group delay spreads pulses and may distort communication symbols.
Bode magnitude uses:
20log10|H(jω)|so cascade gains become sums in dB.
3.5. Resampling, Interpolation, and Sample-Rate Conversion
Changing the sampling rate of a discrete signal is not merely dropping indices or inserting zeros. After upsampling, an interpolation filter suppresses spectral images. Before downsampling, an anti-alias filter suppresses content above the new Nyquist limit.
For a rational conversion ratio
L / M,the conceptual chain is upsample by L, apply an appropriate low-pass filter, and downsample by M. Polyphase implementations avoid many unnecessary operations.
This appears in audio sample-rate conversion, sensor fusion, and alignment of streams with different time bases. General interpolation theory belongs to Numerical Analysis; here the focus is band-limited signals and aliasing.
3.6. Sampling, Aliasing, and Windowing
Sampling theory is more than choosing a rate above twice one frequency. Real signals are rarely perfectly band-limited, so an anti-alias filter must attenuate energy outside the intended band before the ADC.
An FFT of a finite record also includes the spectral effect of the window. Signals that do not align with FFT bins exhibit leakage, and window choice trades main-lobe width against sidelobe suppression.
Amplitude and noise measurements should account for coherent gain and equivalent noise bandwidth so the spectrum becomes a quantitative instrument rather than just a plot.
3.7. Scale and Sampling Conditions in Signal Analysis
An FFT plot is incomplete without record length and sample rate. Frequency resolution, window choice, zero-padding, and amplitude scaling all affect interpretation. Zero-padding does not add spectral information; it only samples the displayed frequency axis more densely.
For LTI systems, time- and frequency-domain solutions can cross-check each other. Convolution and transfer-function methods should agree under the same assumptions.
Physical measurement also includes anti-alias filtering and sensor bandwidth. Checking the Nyquist condition only from the digital sample rate ignores the analog content reaching the ADC.
Testing aliasing before choosing the sample rate. A 700 Hz sinusoid sampled at f_s = 1000 samples/s belongs to the same discrete-frequency equivalence class as a 300 Hz component at the sampling instants. The apparent alias frequency is |700 − 1000| = 300 Hz; phase and sign conventions depend on the sample origin. If the highest analog frequency of interest is 700 Hz, unique ideal reconstruction of a band-limited signal requires f_s > 1400 samples/s; a practical design leaves additional margin for the analog anti-alias filter transition band. A digital filter applied after sampling cannot reliably recover a 700 Hz component that has already aliased. The acquisition procedure must therefore establish analog bandwidth, filtering requirements, and sample rate in that order.
Unit 4: Laplace, Z Transform, and State Space
4.1. Laplace Transform
X(s)=∫x(t)e^(-st)dt
s=σ+jωextends Fourier analysis with exponential weighting and a region of convergence. The ROC is part of the transform because the same algebraic expression can represent different time-domain signals under different ROCs.
For a causal rational system, the ROC lies to the right of the rightmost pole; stability requires the jω axis to lie inside the ROC. A causal stable continuous-time rational system therefore has all poles in the left half-plane.
The system function:
H(s)=Y(s)/X(s)comes directly from the differential equation under initial rest. The unilateral transform includes initial conditions:
L{dx/dt}=sX(s)-x(0-)4.2. Z Transform
The discrete-time counterpart is:
X(z)=Σx[n]z^(-n)and the unit circle corresponds to the DTFT.
For a causal rational discrete-time system, the ROC lies outside the outermost pole; stability requires the unit circle to lie inside the ROC. A causal stable system therefore has all poles inside the unit circle.
Difference equations produce:
H(z)=B(z)/A(z)and poles near the unit circle create longer memory and sharper frequency selectivity.
4.3. Numerical Realization
A high-order IIR implemented as one polynomial is numerically fragile; cascaded second-order sections are usually more robust:
H(z)=H1(z)H2(z)...Hm(z)Overflow, rounding, coefficient quantization, and limit cycles are system effects in finite-precision implementations. Direct, cascade, and parallel realizations of the same theoretical transfer function can behave differently.
4.4. State Space
A transfer function describes input-output behavior; state space exposes internal memory.
Continuous time:
x_dot=Ax+Bu
y=Cx+DuDiscrete time:
x[n+1]=Ax[n]+Bu[n]
y[n]=Cx[n]+Du[n]The state is the smallest internal information set needed to predict future behavior given future input. MIMO systems, controllability, and observability are naturally expressed in this form.
4.5. Controllability, Observability, and the Bridge to Kalman Filtering
For a state-space model
x_(k+1) = A x_k + B u_k
y_k = C x_k + D u_k,controllability asks whether suitable inputs can drive the state through the required directions, while observability asks whether the internal state can be distinguished from output history.
These are relevant not only to control but also to state estimation. In a linear-Gaussian setting, the Kalman filter recursively combines model prediction with measurement information using uncertainty covariances.
The goal here is not to derive every Kalman equation. It is to place the filter in the correct context of state space, noise models, and observability rather than treating it as a generic smoothing routine. Its statistical interpretation connects to conditional Gaussian models in Probability and Statistics.
4.6. Uncertainty and Reduced Representations in State Space
A signal-processing chain may permit many measurement channels to be explained by a smaller latent state vector. The observation matrix C and state-transition matrix A connect physical state variables to measured outputs. A low-rank representation should not be selected solely by the fraction of sample variance it explains: observability and low-energy modes that matter to control may also need preservation. PCA's maximum-variance directions are not necessarily sufficient for dynamic state estimation.
A linear model with stochastic process and measurement noise can be written as x_(k+1) = A x_k + w_k and y_k = C x_k + v_k. Covariance assumptions for w_k and v_k inform the Kalman-filter updates. If these assumptions drift over time or observations have heavy-tailed errors, a filter with low average squared error may still suffer occasional large deviations. Validating the stochastic state model answers a different question from validating a spectral filter's frequency response.
When multiple sensor time series are collected together, common latent factors may be estimated. Good numerical fit of a factor model does not establish that its factors are unique physical mechanisms; transformation and scaling ambiguities must remain explicit. Shared frequency-domain components should also be checked for consistency with time-domain delays, phases and measurement timestamps.
Unit 5: Communications, Feedback, and Real Processing Chains
5.1. Modulation and Communications
Amplitude modulation follows:
x(t)cos(ωct)
<->
(1/2)[X(j(ω-ωc))+X(j(ω+ωc))]Coherent detection needs carrier phase; envelope detection simplifies the receiver at the cost of carrier power. Single-sideband transmission carries the same information in half the DSB bandwidth.
Frequency modulation places information in instantaneous frequency and trades more bandwidth for improved tolerance to amplitude noise.
PAM carries symbol amplitudes, while QAM uses two quadrature carrier components. Band limitation causes intersymbol interference; raised-cosine and root-raised-cosine pulse shaping balance bandwidth and timing robustness.
OFDM divides data among many orthogonal subcarriers implemented efficiently with the DFT/FFT. A cyclic prefix simplifies multipath equalization, while peak-to-average power ratio, carrier-frequency error, and prefix overhead remain its main costs.
5.2. Matched Filter
For a known waveform in white noise, the LTI filter that maximizes SNR at the chosen sampling instant is the matched filter; its impulse response is proportional to the time-reversed conjugate of the target waveform. Radar, sonar, digital communications, and correlation-based detection all use this result.
5.3. Feedback
Negative feedback gives:
Q=H1/(1+H1H2)with closed-loop poles satisfying 1+H1H2=0.
Feedback can reduce sensitivity and disturbances, improve tracking, and stabilize an unstable plant; unfavorable phase can also destabilize an otherwise stable system.
Root locus shows pole motion with gain. The Nyquist criterion relates encirclements of -1 to closed-loop stability. Gain and phase margins measure distance to instability; pure delay changes no magnitude but consumes phase margin.
5.4. Real Processing Chains
Separate theory topics merge in real systems:
physical source
-> analog front end
-> anti-alias filter
-> ADC
-> buffer
-> digital filter
-> transform / feature
-> decision / coding
-> transmission / storageFor real-time work, latency, block size, computational cost, memory access, numerical precision, clock drift, and sample loss matter alongside mathematical correctness. Longer FFTs refine frequency spacing but increase block latency; longer FIR filters sharpen transitions but increase computation and group delay.
5.5. One Framework
time domain transform domain
------------------ ----------------
convolution multiplication
derivative/difference frequency weighting
shift phase factor
exponential input scalar gain
stability poles + ROCFourier series decomposes periodic signals into harmonics, Fourier transform describes aperiodic spectra, Laplace adds the continuous-time ROC, and the z transform adds the discrete-time ROC.
The continuous-time stability boundary is the jω axis and the discrete-time boundary is the unit circle; sampling connects their frequencies through ω=ΩT.
5.6. Key Distinctions
- Linearity requires zero input to produce zero output.
- Real-time physical operation requires causality; offline processing may be noncausal.
- BIBO stability requires an absolutely integrable or summable impulse response.
- A discrete sinusoid is periodic only when normalized frequency is rational.
- DTFT has a continuous periodic frequency variable; DFT has finitely many frequency samples.
- FFT is a DFT algorithm, not a new transform.
- Zero padding does not create true frequency resolution.
- Sampling discretizes time; quantization discretizes amplitude.
- The Nyquist rate is a lower bound, not the chosen sampling rate.
- Aliasing is irreversible information loss.
- Decimation requires prior band limitation.
- FIR is finite and nonrecursive; IIR requires pole-stability analysis.
- Linear phase is not zero phase.
- Laplace and z transforms are incomplete without their ROCs.
- Causal stable continuous-time poles lie in the left half-plane; discrete-time poles lie inside the unit circle.
- Finite-precision realizations of the same transfer function can behave differently.
- Open-loop stability alone does not determine closed-loop stability.
5.7. The Interaction Between Signal Processing and Artificial Intelligence
The relationship between signal processing and AI is bidirectional. Signal processing turns a physical measurement into a sampled, meaningful representation; learning methods use those representations for classification, estimation, separation, or reconstruction. Learned front ends can replace some hand-designed features, but they do not remove sampling theory or measurement physics.
5.8. From Physical Signal to Model Input
Audio, vibration, and communication signals must be measured and sampled. If the Nyquist condition is violated, aliasing folds distinct frequency components together. A larger neural network cannot deterministically reconstruct information that was never sampled correctly.
physical signal
↓
sampling
↓
digital representation
↓
time / frequency / time-frequency transform
↓
learning model5.9. Fourier and Time-Frequency Representation
Speech and many acoustic events are nonstationary. A single Fourier transform of an entire recording hides temporal evolution. STFT exposes local spectra through windows. Window length and hop size trade time resolution against frequency resolution.
If a model consumes spectrograms, sample rate, window, hop, scaling, and normalization are part of the model contract. Training with one front end and serving with another can create distribution shift even when tensor dimensions match.
5.10. Hand-Designed and Learned Features
Classical speech systems often use MFCCs, filter-bank energies, or other engineered features. Modern neural systems may learn representations from waveforms or lower-level features. The distinction is:
classical: signal → designed feature → classifier
learned: signal → learned representation → decisionHybrid designs are common; log-Mel features can encode useful signal knowledge while later representations are learned.
5.11. Two Meanings of Convolution
For an LTI system:
y[n] = Σ x[k] h[n-k]where h is the impulse response. A convolutional neural network uses a similar sliding operation, but a learned kernel is not necessarily a physical impulse response. The operator is related; the modeling interpretation is different.
5.12. Noise, Augmentation, and Robustness
Physical noise and data augmentation are not the same. Augmentation is an imposed training distribution. If it does not represent field conditions, it can optimize the model for the wrong problem. SNR, clipping, reverberation, and channel effects should be tied to the actual operating environment.
5.13. Decision Continuity in Voice Activity Detection
Voice activity detection is broader than labeling every frame independently as speech or silence. Short energy drops, background noise, and transient impulses can make frame-level decisions oscillate. The frame score and the final temporal segment boundary are therefore different objects.
A VAD pipeline can combine short-time energy or spectral-band features with an adaptive noise floor and a speech model. Noise estimates should be updated only from frames that are sufficiently likely to be non-speech; contaminating the noise model with speech can move the decision boundary in the wrong direction.
Temporal stability can be improved through mechanisms such as:
- consecutive support before declaring speech,
- short tolerance before declaring the end of speech,
- hangover or hysteresis,
- state reset after sustained silence.
These mechanisms reduce fragmentation but can increase end-of-segment latency. VAD tuning is therefore a trade-off among latency, fragmentation, and false activation, not a single accuracy number.
For compressed telephony audio or mixed codec representations, signal processing should also start from a well-defined sample representation such as PCM. Codec decoding, sample rate, and channel layout are part of the input contract for later feature extraction; a learned model does not remove those physical and representational assumptions.
5.14. Contributions of Statistical Learning to Signal Processing
Denoising, source separation, speech recognition, acoustic-event detection, and channel estimation can benefit from learned high-dimensional mappings. Yet the output of a learned enhancement model is an estimate, not the original physical signal. In forensic or measurement contexts that distinction is essential.
Evaluation should therefore remain layered:
measurement → SNR / clipping / sampling
signal transform → spectral or waveform error
model task → WER / accuracy / F1 / task metric
system → latency / RTF / throughputImprovement at one layer does not guarantee improvement at another.
Signal processing is not merely preprocessing for AI, and AI is not an automatic replacement for DSP. Signal processing determines what information is represented from the physical world; learning determines what task-relevant mapping can be estimated from that representation.
5.15. Evaluation Layers
No single machine-learning metric is sufficient. A signal-processing system may be measured at multiple levels:
measurement quality → SNR / clipping / sampling
signal transform → spectral or time-domain error
model task → WER / accuracy / F1 / task metric
system → latency / RTF / throughputImprovement at one level does not imply improvement at another. A perceptually cleaner recording, for example, need not yield a lower ASR word error rate.
Signal processing is not merely preprocessing before learning, and learning is not an automatic replacement for digital signal processing. Physical measurement and representation determine what information is encoded, while a learned model determines which decisions can be inferred from that representation. Reliability requires validation across both layers.
5.16. Choosing the Representation Domain for the Problem
The transforms used in Signals and Systems express the same signal in different mathematical languages. The time domain shows when samples or events occur and with what amplitude; the frequency domain makes the spectral components of variation explicit. For a linear time-invariant system, convolution in time becomes multiplication in frequency. A long convolution calculation can therefore become the simpler relation Y(f) = X(f)H(f). If the question concerns switching time, delay, or the temporal structure of an impulse response, the time domain may instead be the clearer representation.
A Fourier series represents a periodic signal by discrete harmonics, whereas the Fourier transform provides a continuous-frequency representation for signals that need not be periodic. The Laplace transform extends frequency reasoning into the complex plane so that growth and decay are represented explicitly; its region of convergence is therefore part of the mathematical meaning rather than a bookkeeping detail. For discrete-time systems, the z-transform plays a comparable role, and pole locations relative to the unit circle connect directly to stability.
The most important sampling distinction is that a high sampling rate cannot repair every mistake after the fact. The Nyquist condition is a reconstruction condition for a signal assumed to be band-limited before sampling. Once distinct spectral components have folded onto one another through aliasing, a later digital filter cannot in general recover the lost distinction.
Before selecting a transform, identify whether the signal is continuous or discrete, periodic or aperiodic, whether the system satisfies the LTI assumptions, and which information the problem asks you to preserve. Fourier, Laplace, and z-domain methods are then not competing formulas but complementary representations of system behaviour under different assumptions.
5.17. Outliers and Robust Feature Extraction in Signal Measurement
Robust statistics asks how much an estimator degrades under small departures from an idealized model. For signal measurements affected by impulsive noise, the median and MAD are often more stable than the mean; influence functions and breakdown points further characterize sensitivity to isolated impulses and to a growing fraction of contaminated samples. In heavy-tailed or contaminated data, decisions based only on the mean and standard deviation can be misleading.
- Show how mean-based features degrade under impulsive noise.
- State which corruption modes benefit from median-based filtering or summaries.
- Preserve the distinction between suppressing anomalies and deleting genuine transients.
5.18. Filter Delay Belongs in the Control-Loop Model
Suppose motor speed is sampled at 1 kHz and processed by a causal five-sample moving average. The symmetric five-tap FIR filter has a constant group delay of (5-1)/2 = 2 samples, or 2 ms. Even if computation itself is nearly instantaneous, the estimate represents an earlier state of motion. Its response H(e^jω) = (1/5)∑(k=0..4)exp(-jωk) also shows that different frequency components are not retained equally.
At a control-loop crossover frequency of 40 Hz, this delay alone contributes approximately 360° × 40 × 0.002 = 28.8° of phase lag. Buffering, communications, and actuation introduce additional delay. A smoother-looking speed trace can therefore come at the expense of phase margin. Measurements should compare both signal-noise behavior and the resulting open-loop phase response. Extending the averaging window is not automatically an improvement in closed-loop control.
Unit 6: Signal-Source Estimation and Multi-Sensor Analysis
6.1. Observation Model for Electromagnetic Signals
A sampled receiver signal can be represented as x_m[n]=(h_m*s)[n]+v_m[n], combining source signal, propagation channel, receiver response, and noise. Receiver position, antenna characteristics, sampling clock, and channel delay are part of the measurement model. Signal detection, content decoding, and source localization are distinct inverse problems.
Narrowband and wideband models do not share identical assumptions. Multipath propagation, co-channel transmissions, and oscillator offsets can bias source estimation. Neither source identity nor exact location follows from a single power or frequency measurement without acquisition context.
6.2. Angle of Arrival and Bearing Intersection
Angle of Arrival (AoA) localization relates bearing estimates acquired at different receiver positions. In a two-dimensional plane, a candidate bearing line can be written as r_m+λ_m d_m, where r_m is receiver position, d_m a unit direction vector, and λ_m a range parameter. Noisy bearings rarely intersect at a single point; weighted least-squares methods can estimate a location.
Closely spaced receivers or nearly parallel bearing lines yield poor geometric conditioning. Angular error, antenna calibration, coordinate transforms, and multipath enlarge uncertainty. An estimated location must be reported with an uncertainty region and model assumptions rather than as an exact coordinate.
6.3. Antenna Arrays and Subspace Direction Estimation
A narrowband antenna-array model is x(t)=a(θ)s(t)+n(t), where a(θ) denotes the angle-dependent steering vector. Decomposing the covariance matrix R=E[xxᴴ] into signal and noise subspaces provides the basis for high-resolution AoA methods such as MUSIC. The method requires appropriate array calibration, source-count assumptions, sufficient samples, and an adequate signal model.
Coherent multipath, element gain/phase mismatch, and violations of narrowband assumptions can invalidate estimates. Spatial resolution depends on aperture, geometry, SNR, and sample availability rather than antenna count alone.
6.4. Time-Difference Measurements and Multi-Receiver Association
For synchronized receivers, a time difference of arrival Δt_ij=t_i−t_j constrains the source distance difference to approximately cΔt_ij, where c is propagation speed in the medium. Multiple receivers produce hyperbolic localization constraints. Common-clock calibration, signal bandwidth, and correlation-peak resolvability are essential prerequisites.
With receiver clock offsets b_i, the measured difference is Δt_measured=Δt_true+b_i−b_j. Reflections and non-line-of-sight propagation degrade accuracy. Phase delay, group delay, carrier offsets, and receiver buffering must not be confused with source movement.
6.5. Spectral and Temporal Signal Characteristics
Carrier frequency, bandwidth, transmission duration, duty cycle, power spectral density, burst intervals, and repetition rate describe transmissions without content decoding. These measurements do not independently identify a person or transmitter. Spectral-window choice and observation length affect both time and frequency resolution.
Feature fusion must account for different sampling characteristics, missing observations, and multiple receivers recording the same emission. Classification remains an inference layer distinct from direct physical measurement.
6.6. Multi-Sensor Fusion and Estimation Validity
Fusion across sensors requires common time references, coordinate systems, uncertainty models, and source-association criteria. Weighted least squares or probabilistic fusion may be appropriate for independent observations; correlated errors can lead to overconfident uncertainty estimates if independence is assumed incorrectly.
Validation should include known source positions, controlled multipath, low SNR, and clock-misalignment conditions. High-percentile errors and failed-estimation rates must accompany mean position error. Credible conclusions depend on validating both measurement and estimation models.
6.7. Spectral Representations, Cepstrum, and Time-Frequency Resolution
In the short-time Fourier transform, window duration trades frequency resolution against temporal localization. A short window localizes transients, whereas a longer window can separate nearby frequency components more effectively. The real cepstrum, derived from an inverse transform of the log-magnitude spectrum, can help examine source and filter structure across quefrency regions, but does not itself perform physical source separation.
An MFCC pipeline built from mel filter banks and a discrete cosine transform is related to, but not identical with, the real cepstrum. Coefficient count, frame shift, pre-emphasis, and normalization affect model input. Claims about audio-processing performance require the feature parameters, sampling frequency, and evaluation conditions.
6.8. Multiscale Analysis of Transient Events
Shortening an STFT window improves temporal localisation at the expense of frequency resolution. Wavelet decompositions can examine transients at several scales. Broadband impulsive energy does not establish corruption: a speech attack or physical impact may have the same signature. Peak amplitude, timing and spectral structure should be checked after filtering. Event detection versus correction is developed in Anomaly Detection.
Related Courses
Sampling, Fourier analysis, convolution and filtering are applied in Audio Processing and Forensic Audio Engineering and Image Processing and Computer Vision Engineering. System dynamics and feedback are studied in Automatic Control, while physical circuit models are examined in Circuit Theory.
References
- Alan V. Oppenheim; Ronald W. Schafer. (2004). From Frequency to Quefrency: A History of the Cepstrum. IEEE Signal Processing Magazine, 21(5), 95-106. doi:10.1109/MSP.2004.1328092
- Alec Radford; Jong Wook Kim; Tao Xu; Greg Brockman; Christine McLeavey; Ilya Sutskever. (2022). Robust Speech Recognition via Large-Scale Weak Supervision. arXiv. doi:10.48550/arXiv.2212.04356
- Alex Graves, Abdel-rahman Mohamed, Geoffrey Hinton. “Speech Recognition with Deep Recurrent Neural Networks.” ICASSP, 2013. https://doi.org/10.1109/ICASSP.2013.6638947
- Arm. CMSIS-DSP: Finite Impulse Response (FIR) Filters. https://arm-software.github.io/CMSIS-DSP/main/group__FIR.html
- Billingsley, J. Mechatronics Experiments with Machine Vision. Springer Nature Switzerland, 2026. DOI: https://doi.org/10.1007/978-3-032-11218-7.
- Cooley, J. W., & Tukey, J. W. “An Algorithm for the Machine Calculation of Complex Fourier Series.” Mathematics of Computation, 19(90), 1965. https://doi.org/10.1090/S0025-5718-1965-0178586-1
- Franklin, G. F., Powell, J. D., & Emami-Naeini, A. Feedback Control of Dynamic Systems, 8th ed. Pearson, 2019.
- Geoffrey Hinton et al. “Deep Neural Networks for Acoustic Modeling in Speech Recognition.” IEEE Signal Processing Magazine, 29(6), 2012. https://doi.org/10.1109/MSP.2012.2205597
- Harris, F. J. “On the Use of Windows for Harmonic Analysis with the Discrete Fourier Transform.” Proceedings of the IEEE, 66(1), 1978. https://doi.org/10.1109/PROC.1978.10837
- Haykin, S., & Van Veen, B. Signals and Systems, 2nd ed. Wiley, 2003.
- Kay, S. M. Fundamentals of Statistical Signal Processing: Estimation Theory. Prentice Hall, 1993.
- Lathi, B. P. Linear Systems and Signals, 2nd ed. Oxford University Press, 2005.
- Lyons, R. G. Understanding Digital Signal Processing, 3rd ed. Pearson, 2011.
- Matthews, P. SIGINT: The Secret History of Signals Intelligence 1914–45. The History Press, 2013.
- Mor Harchol-Balter. Performance Modeling and Design of Computer Systems: Queueing Theory in Action. Cambridge University Press, 2013.
- Müller, M. Audio Identification. Fundamentals of Music Processing Notebooks, Friedrich-Alexander-Universität Erlangen-Nürnberg. https://www.audiolabs-erlangen.de/resources/MIR/FMP/C7/C7S1_AudioIdentification.html
- Necsulescu, D. Advanced Mechatronics: Monitoring and Control of Spatially Distributed Systems. World Scientific, 2009. ISBN 978-981-277-181-0.
- Oppenheim, A. V., & Schafer, R. W. Discrete-Time Signal Processing, 3rd ed. Pearson, 2010.
- Oppenheim, A. V., Willsky, A. S., & Nawab, S. H. Signals and Systems, 2nd ed. Prentice Hall, 1997.
- Peter J. Huber; Elvezio M. Ronchetti. Robust Statistics. 2nd ed. Wiley, 2009.
- Proakis, J. G., & Manolakis, D. G. Digital Signal Processing: Principles, Algorithms, and Applications, 4th ed. Pearson, 2007.
- Proakis, J. G., & Salehi, M. Digital Communications, 5th ed. McGraw-Hill, 2007.
- SciPy. scipy.signal.convolve. https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.convolve.html
- Shannon, C. E. “Communication in the Presence of Noise.” Proceedings of the IRE, 37(1), 1949. https://doi.org/10.1109/JRPROC.1949.232969
- Steven Davis; Paul Mermelstein. (1980). Comparison of Parametric Representations for Monosyllabic Word Recognition in Continuously Spoken Sentences. IEEE Transactions on Acoustics, Speech, and Signal Processing, 28(4), 357-366. doi:10.1109/TASSP.1980.1163420
- Steven W. Smith. The Scientist and Engineer's Guide to Digital Signal Processing, chapter 18: FFT Convolution. https://www.dspguide.com/CH18.PDF
- Van Trees, H. L. Optimum Array Processing: Part IV of Detection, Estimation, and Modulation Theory. Wiley, 2002.
- Walter W. Piegorsch, Richard A. Levine, Hao Helen Zhang, Thomas C. M. Lee (eds.). Computational Statistics in Data Science. Wiley, 2022. ISBN 9781119561071.
- Wang, A. L.-C. "An Industrial-Strength Audio Search Algorithm." ISMIR 2003, 7–13, 2003. https://www.princeton.edu/~cuff/ele301/files/Wang03-shazam.pdf
- Welch, P. D. “The Use of Fast Fourier Transform for the Estimation of Power Spectra.” IEEE Transactions on Audio and Electroacoustics, 15(2), 1967. https://doi.org/10.1109/TAU.1967.1161901
- West, N. The SIGINT Secrets: The Signals Intelligence War, 1900 to Today. William Morrow, 1988.
- Stéphane Mallat. A Wavelet Tour of Signal Processing: The Sparse Way, 3rd ed.. Academic Press, 2009. https://doi.org/10.1016/B978-0-12-374370-1.X0001-8