Spectrograms in MATLAB - Deep Dive

Overview - Spectrograms

What is it?

A spectrogram is a visual way to see how the frequencies in a sound or signal change over time. It breaks the signal into small pieces and shows which frequencies are strong in each piece. This helps us understand sounds or signals that change, like music or speech. The result is a colorful image where one axis is time, another is frequency, and colors show strength.

Why it matters

Without spectrograms, it would be very hard to analyze signals that change over time, like spoken words or bird songs. They let us see hidden patterns and details that simple graphs miss. This helps in many areas like music, medicine, and engineering to detect problems or understand complex sounds. Without them, many modern technologies like speech recognition or noise filtering would be much less effective.

Where it fits

Before learning spectrograms, you should understand basic signals and how to use the Fourier transform to find frequencies in a signal. After spectrograms, you can explore advanced signal processing techniques like wavelets or machine learning on time-frequency data.

Mental Model

Core Idea

A spectrogram shows how the strength of different frequencies in a signal changes over time by slicing the signal into small time windows and analyzing each separately.

Think of it like...

Imagine listening to a song and writing down which instruments play at each moment. The spectrogram is like a colorful sheet music that shows which notes (frequencies) are loud or soft as time passes.

Time →
Frequency ↓
┌─────────────────────────────┐
│ ███    ██   ████  █         │
│ ██ █  ███  ██ ██ ████       │
│  ███ █ █ █ █  █  █  █       │
│   █  █  █  █   █  █  █      │
│                             │
│ Color = Strength (Amplitude) │
└─────────────────────────────┘

Build-Up - 7 Steps

1

Foundation

Understanding Signals and Frequencies

Concept: Signals can be broken down into frequencies using the Fourier transform.

A signal is a sequence of values over time, like a sound wave. The Fourier transform changes this signal into a list of frequencies and how strong each is. This helps us see what tones make up the sound.

Result

You get a frequency spectrum showing which frequencies are present and their strengths.

Understanding that signals are made of frequencies is the base for seeing how these frequencies change over time.

2

Foundation

Time-Frequency Tradeoff Basics

Concept: Shorter time windows give better time detail but worse frequency detail, and longer windows do the opposite.

When analyzing signals, you cut them into small pieces (windows). If the window is short, you see when things happen better but lose frequency clarity. If it's long, you see frequencies clearly but lose exact timing.

Result

You learn that choosing window size is a balance between time and frequency clarity.

Knowing this tradeoff helps you pick the right window size for your analysis goals.

3

Intermediate

Creating a Spectrogram Step-by-Step

Concept: A spectrogram is made by applying the Fourier transform to many overlapping windows of the signal.

1. Split the signal into overlapping windows. 2. Apply Fourier transform to each window. 3. Calculate the magnitude (strength) of each frequency. 4. Stack these results over time to form a 2D image.

Result

You get a matrix where rows are frequencies, columns are time windows, and values show strength.

Seeing the spectrogram as many small frequency analyses over time clarifies how it captures changing signals.

4

Intermediate

Window Functions and Their Effects

🤔Before reading on: do you think using a rectangular window is always best for spectrograms? Commit to yes or no.

Concept: Window functions shape each segment to reduce edge effects and improve frequency accuracy.

Simply cutting the signal creates sharp edges causing artifacts. Window functions like Hamming or Hann taper the edges smoothly. This reduces false frequencies and improves the spectrogram's clarity.

Result

Spectrograms with proper windows have clearer frequency patterns and fewer artifacts.

Understanding window functions prevents common mistakes that blur or distort frequency information.

5

Intermediate

Overlap and Resolution in Spectrograms

🤔Before reading on: does increasing overlap between windows improve time resolution, frequency resolution, or both? Commit to your answer.

Concept: Overlapping windows increase time resolution without losing frequency resolution.

By overlapping windows, you get more frequent snapshots of the signal. This means smoother changes over time in the spectrogram. The frequency resolution stays the same because window length doesn't change.

Result

Spectrograms look smoother and more detailed in time with overlap.

Knowing how overlap affects resolution helps optimize spectrogram quality for your needs.

6

Advanced

Interpreting Spectrogram Patterns

🤔Before reading on: do you think a horizontal line in a spectrogram means a constant frequency or a changing frequency? Commit to your answer.

Concept: Different shapes in a spectrogram correspond to different signal behaviors like steady tones or chirps.

A horizontal line means a steady frequency over time. A rising or falling line shows frequency changing (like a siren). Vertical lines indicate sudden bursts or clicks. Understanding these helps identify sounds or events.

Result

You can read spectrograms to recognize signal features and events.

Interpreting patterns turns spectrograms from pictures into meaningful information.

7

Expert

Limitations and Advanced Techniques

🤔Before reading on: do you think spectrograms can perfectly represent all signals without any loss? Commit to yes or no.

Concept: Spectrograms have limits due to time-frequency tradeoff and windowing; advanced methods like wavelets can improve analysis.

Because of fixed window sizes, spectrograms can't perfectly show all details. Fast changes may blur, and close frequencies may merge. Wavelet transforms adapt window size to frequency, giving better detail. Also, choosing parameters wrong can hide important info.

Result

You understand when spectrograms work well and when to use other tools.

Knowing spectrogram limits prevents misinterpretation and guides better analysis choices.

Under the Hood

Internally, the spectrogram divides the signal into overlapping segments. Each segment is multiplied by a window function to reduce edge effects. Then, the Fourier transform converts each windowed segment from time domain to frequency domain, producing complex numbers representing amplitude and phase. The magnitude of these numbers forms the intensity values in the spectrogram. Stacking these magnitudes over time creates a 2D time-frequency representation.

Why designed this way?

Spectrograms were designed to analyze non-stationary signals whose frequency content changes over time. The short-time Fourier transform (STFT) approach balances time and frequency resolution by using windows. Alternatives like wavelets exist but STFT is simple, efficient, and widely supported, making it a practical choice for many applications.

Signal → [Window 1] → Fourier Transform → Magnitude → Spectrogram Column 1
       → [Window 2] → Fourier Transform → Magnitude → Spectrogram Column 2
       → [Window 3] → Fourier Transform → Magnitude → Spectrogram Column 3
       ...

Each window overlaps with neighbors to capture smooth changes over time.

Myth Busters - 4 Common Misconceptions

Quick: Does a brighter color in a spectrogram always mean a louder sound? Commit to yes or no.

Common Belief: Brighter colors always mean louder sounds in the original signal.

Reality: Brighter colors mean stronger frequency components in that time window, but not necessarily louder overall sound because perception depends on many factors.

Why it matters: Misinterpreting brightness can lead to wrong conclusions about sound intensity or importance.

Quick: Is a spectrogram just a fancy picture of the signal waveform? Commit to yes or no.

Common Belief: A spectrogram is just a visual of the signal's waveform over time.

Reality: A spectrogram shows frequency content over time, which is very different from the raw waveform shape.

Why it matters: Confusing waveform and spectrogram leads to misunderstanding what information each provides.

Quick: Does increasing window length always improve spectrogram quality? Commit to yes or no.

Common Belief: Longer windows always make spectrograms better by giving clearer frequency info.

Reality: Longer windows improve frequency resolution but reduce time resolution, possibly missing fast changes.

Why it matters:

Quick: Can spectrograms perfectly separate two frequencies that are very close together? Commit to yes or no.

Common Belief: Spectrograms can always distinguish any two frequencies no matter how close.

Reality: Spectrogram frequency resolution is limited by window size; very close frequencies may blur together.

Why it matters: Overestimating resolution can cause missed details or false interpretations.

Expert Zone

1

The choice of window function affects sidelobe levels, which can cause leakage of strong frequencies into nearby bins, misleading interpretation.

2

Overlap percentage impacts computational cost and smoothness; too much overlap increases processing time without much gain.

3

Phase information in the Fourier transform is often ignored in spectrograms but can be crucial for advanced signal reconstruction or analysis.

When NOT to use

Spectrograms are not ideal for signals with very fast transient events or when adaptive time-frequency resolution is needed. Alternatives like wavelet transforms or reassigned spectrograms provide better localization. For stationary signals, simple Fourier analysis suffices.

Production Patterns

In real-world systems, spectrograms are used for voice activity detection, music genre classification, and fault diagnosis in machines. They are often combined with machine learning models that take spectrogram images as input for automated pattern recognition.

Connections

Fourier Transform

Spectrograms build directly on the Fourier transform by applying it to short windows of the signal.

Understanding Fourier transform is essential to grasp how spectrograms reveal frequency content over time.

Wavelet Transform

Wavelets provide an alternative time-frequency analysis with adaptive window sizes, improving on spectrogram limitations.

Knowing spectrogram limits helps appreciate why wavelets are used for signals with varying frequency details.

Music Sheet Notation

Both spectrograms and music sheets represent sound frequencies over time but in different formats and purposes.

Recognizing this connection helps relate abstract signal analysis to familiar musical concepts.

Common Pitfalls

#1 Using a rectangular window without tapering causes spectral leakage.

Wrong approach:

window = ones(window_length,1);
[S,F,T] = spectrogram(signal, window, overlap, nfft, fs);

Correct approach:

window = hamming(window_length);
[S,F,T] = spectrogram(signal, window, overlap, nfft, fs);

Root cause:

Not applying a window function leads to sharp edges in segments, causing frequency smearing.

#2 Choosing window length too long for fast-changing signals.

Wrong approach:

window_length = length(signal) / 2; % very long window
[S,F,T] = spectrogram(signal, window_length, overlap, nfft, fs);

Correct approach:

window_length = appropriate_short_length; % e.g., 256 samples
[S,F,T] = spectrogram(signal, window_length, overlap, nfft, fs);

Root cause:

Ignoring time-frequency tradeoff causes loss of temporal detail.

#3 Not overlapping windows, resulting in choppy spectrogram.

Wrong approach:

[S,F,T] = spectrogram(signal, window, 0, nfft, fs); % zero overlap

Correct approach:

[S,F,T] = spectrogram(signal, window, floor(length(window)*0.5), nfft, fs); % 50% overlap

Root cause:

No overlap reduces smoothness and time resolution.

Key Takeaways

Spectrograms visualize how frequencies in a signal change over time by applying Fourier transform to short, overlapping windows.

Choosing window size and type balances time and frequency resolution, which is crucial for accurate analysis.

Window functions reduce artifacts and improve clarity by smoothing edges of signal segments.

Interpreting spectrogram patterns helps identify steady tones, frequency changes, and transient events in signals.

Spectrograms have limits due to fixed window sizes; advanced methods like wavelets can overcome some of these.

Practice

1. What does a spectrogram primarily display in signal processing?

easy

A. The total energy of a signal

B. How frequencies change over time in a signal

C. The amplitude of a signal at a single point

D. The phase shift of a signal

Solution

1. Step 1: Understand the purpose of a spectrogram

   A spectrogram is a visual tool that shows how the frequency content of a signal varies over time.

2. Step 2: Compare options with the definition

   Only How frequencies change over time in a signal correctly describes this as frequency changes over time, while others describe different signal properties.

3. Final Answer:

   How frequencies change over time in a signal -> Option B

4. Quick Check:

   Spectrogram = frequency vs time [OK]

Hint: Spectrogram = frequency changes over time [OK]

Common Mistakes:

• Confusing amplitude with frequency
• Thinking spectrogram shows total energy
• Mixing phase shift with frequency content

2. Which of the following is the correct syntax to create a spectrogram in MATLAB for signal x with window size 256 and overlap 128?

easy

A. spectrogram(x,256,128)

B. spectrogram(x,128,256)

C. spectrogram(256,x,128)

D. spectrogram(x,128)

Solution

1. Step 1: Recall MATLAB spectrogram syntax

   The syntax is spectrogram(signal,window,overlap), so the window size comes before overlap.

2. Step 2: Match parameters to syntax

   spectrogram(x,256,128) correctly uses spectrogram(x,256,128) where 256 is window size and 128 is overlap.

3. Final Answer:

   spectrogram(x,256,128) -> Option A

4. Quick Check:

   Correct order: signal, window, overlap [OK]

Hint: Remember: spectrogram(signal, window, overlap) [OK]

Common Mistakes:

• Swapping window and overlap values
• Passing window size as first argument
• Omitting required parameters

3. What will be the output size of the matrix S when running [S,F,T] = spectrogram(x,128,64,256,1000); if x is a 1-second signal sampled at 1000 Hz?

medium

A. 256 rows by 15 columns

B. 128 rows by 16 columns

C. 129 rows by 14 columns

D. 129 rows by 16 columns

Solution

1. Step 1: Calculate frequency bins (rows)

   Frequency bins = nfft/2 + 1 = 256/2 + 1 = 129 rows.

2. Step 2: Calculate time bins (columns)

   Signal length = 1000 samples, window=128, overlap=64, step = 128-64=64 samples.
   Number of segments = floor((1000-128)/64) + 1 = floor(872/64) + 1 = 13 + 1 = 14 columns.

3. Final Answer:

   129 rows by 14 columns -> Option C

4. Quick Check:

   Rows = 129, Columns = 14 [OK]

Hint: Rows = nfft/2+1, columns depend on window and overlap [OK]

Common Mistakes:

• Confusing nfft with window size for rows
• Miscounting number of segments
• Ignoring overlap in time bins calculation

4. Given the code snippet:

x = randn(1,1000);
[S,F,T] = spectrogram(x,256,128,512,1000);
imagesc(T,F,abs(S)); axis xy;

The spectrogram plot looks very blurry. What is the most likely cause?

medium

A. Sampling frequency is too low

B. Overlap is too small, causing gaps in data

C. nfft is smaller than window size, causing error

D. Window size is too large, causing poor time resolution

Solution

1. Step 1: Analyze window size effect on resolution

   A large window size (256) smooths over time, reducing time resolution and causing blur.

2. Step 2: Check other parameters

   Overlap is 128 (50%), which is typical and not causing gaps.
   nfft (512) is larger than window (256), which is valid.
   Sampling frequency (1000) is standard.

3. Final Answer:

   Window size is too large, causing poor time resolution -> Option D

4. Quick Check:

   Large window = poor time resolution [OK]

Hint: Large window = better frequency, worse time resolution [OK]

Common Mistakes:

• Confusing overlap with window size effect
• Assuming nfft smaller than window causes blur
• Blaming sampling frequency for blur

5. You want to analyze a signal with rapidly changing frequencies using MATLAB's spectrogram. Which combination of parameters will best capture quick changes without losing frequency detail?

hard

A. Small window size, large overlap, moderate nfft

B. Large window size, large overlap, small nfft

C. Large window size, small overlap, large nfft

D. Small window size, small overlap, large nfft

Solution

1. Step 1: Understand trade-offs in spectrogram parameters

   Small window size improves time resolution to capture rapid changes.
   Large overlap ensures smooth transitions between windows.
   Moderate nfft balances frequency resolution without excessive computation.

2. Step 2: Evaluate options

   Small window size, large overlap, moderate nfft matches these criteria.
   Small window size, small overlap, large nfft has small overlap, causing gaps.
   Options B and C use large windows, reducing time resolution.

3. Final Answer:

   Small window size, large overlap, moderate nfft -> Option A

4. Quick Check:

   Quick changes need small window + large overlap [OK]

Hint: Small window + large overlap = best for fast changes [OK]

Common Mistakes:

• Using large window for fast changes
• Ignoring overlap size
• Choosing too small nfft losing frequency detail