Audio signal processing in MATLAB - Deep Dive

Overview - Audio signal processing

What is it?

Audio signal processing is the way computers handle sounds to analyze, change, or improve them. It involves turning sound waves into numbers that a computer can understand and then using math to work with those numbers. This helps us do things like clean up noisy recordings, recognize speech, or create music effects. It makes sounds easier to use and understand in many devices and applications.

Why it matters

Without audio signal processing, our devices would struggle to understand or improve sounds. Imagine trying to talk to a phone that can't clean up background noise or a music app that can't add effects. This technology makes communication clearer, music richer, and machines smarter about sound. It impacts everyday life from phone calls to hearing aids and voice assistants.

Where it fits

Before learning audio signal processing, you should know basic math like algebra and some programming skills, especially in MATLAB. After this, you can explore advanced topics like speech recognition, music information retrieval, or machine learning with audio data.

Mental Model

Core Idea

Audio signal processing is about turning sound waves into numbers and using math to analyze or change those sounds.

Think of it like...

It's like turning a song into a recipe list of ingredients and steps, so you can change the flavor or fix mistakes before cooking it again.

Sound Wave (Air Vibrations)
      ↓
[Microphone converts]
      ↓
Digital Signal (Numbers)
      ↓
[Processing: filtering, transforming]
      ↓
Modified Signal
      ↓
[Speaker plays]
      ↓
Sound Wave (Modified)

Build-Up - 7 Steps

1

Foundation

Understanding Sound Waves Basics

Concept: Sound is a wave that moves through air, and we can measure its changes over time.

Sound waves are vibrations that travel through air. These vibrations have properties like frequency (how fast they vibrate) and amplitude (how strong they are). Frequency relates to pitch (high or low sound), and amplitude relates to loudness. We can represent these waves as graphs showing how air pressure changes over time.

Result

You can visualize sound as a wave with peaks and valleys representing pressure changes.

Understanding sound as a wave helps you see why we need to convert it into numbers to work with it on a computer.

2

Foundation

Converting Sound to Digital Signals

Concept: Microphones turn sound waves into electrical signals, which computers convert into numbers using sampling.

A microphone captures sound vibrations and changes them into electrical signals. Then, an analog-to-digital converter (ADC) samples these signals many times per second (sampling rate) and records the amplitude as numbers. For example, a common sampling rate is 44,100 samples per second (44.1 kHz), which captures most sounds humans can hear.

Result

You get a sequence of numbers representing the sound wave over time, called a digital audio signal.

Knowing how sound becomes digital data is key to understanding how computers process audio.

3

Intermediate

Basic Audio Signal Processing Techniques

🤔Before reading on: do you think filtering removes all sound or just certain parts? Commit to your answer.

Concept: Filtering lets us remove or keep certain parts of the sound based on frequency, like removing noise or isolating voices.

Filters work by allowing some frequencies to pass and blocking others. For example, a low-pass filter keeps low frequencies and removes high ones, which can reduce hiss noise. MATLAB has built-in functions to create and apply filters to audio signals, changing the sound by emphasizing or reducing parts.

Result

Filtered audio sounds cleaner or focused on certain sounds, like removing background noise.

Understanding filtering helps you control which parts of sound you want to keep or remove, a foundation for many audio tasks.

4

Intermediate

Using Fourier Transform to Analyze Sound

🤔Before reading on: do you think Fourier Transform changes the sound or just shows its parts? Commit to your answer.

Concept: Fourier Transform breaks down a sound into its frequency components, showing what pitches are present and their strength.

The Fourier Transform converts a time-based signal into frequencies. In MATLAB, the fft() function does this. For example, a musical note has a main frequency and harmonics. By looking at the transform, you see peaks at these frequencies. This helps identify sounds or detect patterns.

Result

You get a frequency spectrum showing which frequencies are in the sound and how strong they are.

Knowing how to see sound in frequency terms lets you analyze and manipulate audio more precisely.

5

Intermediate

Applying Time-Frequency Analysis with Spectrograms

🤔Before reading on: do you think a spectrogram shows sound over time, frequency, or both? Commit to your answer.

Concept: A spectrogram shows how frequencies in a sound change over time, combining time and frequency views.

A spectrogram splits the audio into small time windows and applies Fourier Transform to each. This creates a 2D image where one axis is time, the other is frequency, and color shows strength. MATLAB's spectrogram() function creates this. It helps visualize speech, music, or noise patterns changing over time.

Result

You see a colorful image representing how sound frequencies evolve during the recording.

Time-frequency analysis reveals dynamic sound features that simple frequency or time views miss.

6

Advanced

Noise Reduction Using Spectral Subtraction

🤔Before reading on: do you think noise reduction removes all sound or just noise parts? Commit to your answer.

Concept: Spectral subtraction estimates noise in frequency space and subtracts it to clean the audio without losing important sounds.

This method assumes noise is steady and can be estimated from silent parts. It subtracts the noise spectrum from the noisy audio spectrum, then reconstructs the cleaned signal. MATLAB code uses fft, subtraction, and inverse fft. It improves clarity in recordings with background noise.

Result

The output audio has less background noise but keeps the main sounds intact.

Understanding noise reduction in frequency space helps improve audio quality in real-world noisy environments.

7

Expert

Advanced Audio Effects with Digital Filters

🤔Before reading on: do you think digital filters can create effects like echo or reverb? Commit to your answer.

Concept: Digital filters can be designed to create complex audio effects like echo, reverb, or equalization by shaping the signal in time and frequency.

Using MATLAB, you can design filters with specific frequency responses or delays. For example, an echo effect delays the signal and adds it back to the original. Reverb simulates reflections by mixing multiple delayed signals. These effects rely on filter design and signal mixing, controlled by parameters for creativity.

Result

You get audio with added effects that change how it sounds, like making it feel like it's in a big hall.

Knowing how to design and apply digital filters unlocks creative control over sound beyond simple cleaning or analysis.

Under the Hood

Audio signal processing works by sampling continuous sound waves into discrete numbers, then applying mathematical operations like filtering and transforms on these numbers. The computer treats the audio as arrays of values and uses algorithms to change or analyze them. Fourier Transform uses complex math to convert time data into frequency data, revealing hidden patterns. Filters use equations to emphasize or reduce certain frequencies. The processed numbers can be converted back to sound waves for listening.

Why designed this way?

This approach was designed to make sound usable by digital computers, which only understand numbers. Early analog methods were limited and less flexible. Digital processing allows precise, repeatable, and complex operations on sound. The use of Fourier Transform and filters comes from mathematical breakthroughs that let us understand and manipulate signals in frequency terms, which is more intuitive for many audio tasks.

┌───────────────┐      ┌───────────────┐      ┌───────────────┐
│ Sound Waves   │─────▶│ ADC Sampling  │─────▶│ Digital Signal│
└───────────────┘      └───────────────┘      └───────────────┘
                                │                      │
                                ▼                      ▼
                      ┌─────────────────┐      ┌───────────────┐
                      │ Signal Processing│◀────│ Fourier Transform│
                      └─────────────────┘      └───────────────┘
                                │                      │
                                ▼                      ▼
                      ┌───────────────┐      ┌───────────────┐
                      │ Filter/Effects │      │ Inverse FFT   │
                      └───────────────┘      └───────────────┘
                                │                      │
                                ▼                      ▼
                      ┌───────────────┐      ┌───────────────┐
                      │ DAC Conversion│◀────│ Processed Signal│
                      └───────────────┘      └───────────────┘
                                │
                                ▼
                      ┌───────────────┐
                      │ Output Sound  │
                      └───────────────┘

Myth Busters - 4 Common Misconceptions

Quick: Does increasing sampling rate always improve audio quality? Commit to yes or no.

Common Belief: Higher sampling rates always make audio sound better.

Reality: After a certain point (like 44.1 kHz), increasing sampling rate yields little audible improvement and increases data size.

Why it matters: Using unnecessarily high sampling rates wastes storage and processing power without real benefit.

Quick: Does filtering always remove noise without affecting the main sound? Commit to yes or no.

Common Belief: Filters can perfectly remove noise without changing the desired sound.

Reality: Filters often affect parts of the main sound too, causing distortion or loss of quality.

Why it matters: Expecting perfect noise removal can lead to disappointment and poor audio results.

Quick: Is the Fourier Transform reversible without any loss? Commit to yes or no.

Common Belief: Fourier Transform changes the sound and loses information.

Reality: Fourier Transform is mathematically reversible, so no information is lost if done correctly.

Why it matters: Knowing this allows safe frequency domain processing and reconstruction of original signals.

Quick: Does a spectrogram show exact timing of sounds? Commit to yes or no.

Common Belief: Spectrograms show precise timing and frequency details simultaneously.

Reality: Spectrograms trade off time and frequency resolution; you can't have perfect detail in both at once.

Why it matters: Misunderstanding this leads to wrong interpretations of audio features.

Expert Zone

1

Digital audio processing precision depends on bit depth, affecting dynamic range and noise floor, which many beginners overlook.

2

Windowing functions used before Fourier Transform affect spectral leakage and resolution, a subtle but critical detail in analysis.

3

Latency introduced by processing algorithms can impact real-time applications like live audio or hearing aids, requiring careful design.

When NOT to use

Audio signal processing techniques relying on linear assumptions may fail with highly non-linear or distorted sounds; in such cases, machine learning or adaptive filtering methods are better alternatives.

Production Patterns

In real-world systems, audio processing pipelines combine noise reduction, equalization, compression, and effects in stages, often optimized for low latency and resource use, such as in mobile phones or streaming services.

Connections

Image Processing

Both use transforms like Fourier to analyze frequency components, but in different dimensions (1D for audio, 2D for images).

Understanding frequency analysis in audio helps grasp similar concepts in image filtering and enhancement.

Human Hearing Physiology

Audio processing mimics how the ear separates sounds into frequencies and processes them over time.

Knowing ear function guides designing filters and effects that sound natural to humans.

Financial Time Series Analysis

Both analyze signals changing over time and use transforms to find patterns or trends.

Techniques from audio signal processing can inspire methods to detect cycles or anomalies in financial data.

Common Pitfalls

#1 Using too low a sampling rate causing distorted or missing sounds.

Wrong approach:

fs = 8000; % Sampling rate too low for music
soundData = audioread('music.wav');
processed = resample(soundData, fs, originalFs);

Correct approach:

fs = 44100; % Standard sampling rate for audio
soundData = audioread('music.wav');
processed = resample(soundData, fs, originalFs);

Root cause:

Misunderstanding the Nyquist theorem and human hearing range leads to choosing inadequate sampling rates.

#2 Applying a filter without considering its effect on phase causing audio artifacts.

Wrong approach:

filteredAudio = filter(b, a, audioSignal); % Using causal filter directly

Correct approach:

filteredAudio = filtfilt(b, a, audioSignal); % Zero-phase filtering to avoid distortion

Root cause:

Not realizing that causal filters introduce phase shifts that can distort audio timing.

#3 Ignoring windowing effects when performing Fourier Transform leading to spectral leakage.

Wrong approach:

spectrum = fft(audioSignal); % No window applied

Correct approach:

window = hamming(length(audioSignal));
spectrum = fft(audioSignal .* window);

Root cause:

Overlooking the need to reduce edge effects in finite-length signals causes inaccurate frequency analysis.

Key Takeaways

Audio signal processing converts sound waves into numbers so computers can analyze and change sounds.

Sampling rate and bit depth are crucial choices that affect audio quality and data size.

Filters and Fourier Transform are fundamental tools to isolate, analyze, and modify sound frequencies.

Time-frequency analysis like spectrograms reveals how sounds change over time, essential for speech and music.

Advanced techniques like noise reduction and digital effects rely on understanding both time and frequency domains.

Practice

1. What does the MATLAB function audioread do in audio signal processing?

easy

A. It saves audio data to a file.

B. It plays the audio sound through speakers.

C. It reads an audio file and loads the sound data into MATLAB.

D. It filters the audio signal to remove noise.

Solution

1. Step 1: Understand the purpose of audioread

   The function audioread is used to load audio data from a file into MATLAB as a vector or matrix.

2. Step 2: Differentiate from other audio functions

   Functions like sound play audio, audiowrite saves audio, and filter processes signals. audioread only reads data.

3. Final Answer:

   It reads an audio file and loads the sound data into MATLAB. -> Option C

4. Quick Check:

   audioread = read audio data [OK]

Hint: Remember: 'read' means load data, not play or save [OK]

Common Mistakes:

• Confusing audioread with sound
• Thinking it saves audio instead of loading
• Assuming it filters or changes the audio

2. Which of the following is the correct syntax to apply a simple low-pass filter b, a to an audio signal x in MATLAB?

easy

A. y = filter(b, a, x);

B. y = filter(x, b, a);

C. y = filter(a, b, x);

D. y = filter(x, a, b);

Solution

1. Step 1: Recall the filter function syntax

   The MATLAB filter function syntax is y = filter(b, a, x); where b and a are filter coefficients and x is the input signal.

2. Step 2: Identify correct argument order

   Options A, C, and D have the arguments in wrong order, which will cause errors or wrong output.

3. Final Answer:

   y = filter(b, a, x); -> Option A

4. Quick Check:

   filter(b, a, x) = correct syntax [OK]

Hint: Filter syntax is always filter(b, a, x) [OK]

Common Mistakes:

• Swapping filter coefficients order
• Putting signal before coefficients
• Using wrong variable names

3. What will be the output size of y after running this MATLAB code?

x = randn(44100,1); % 1 second noise at 44.1kHz
b = [1 0 0.5]; a = 1;
y = filter(b, a, x);

medium

A. An error due to filter coefficients

B. A vector of size 44102x1

C. A scalar value

D. A vector of size 44100x1

Solution

1. Step 1: Understand input and filter output length

   The input x is a 44100x1 vector. The filter function outputs a vector of the same length as input.

2. Step 2: Check filter coefficients validity

   Coefficients b and a are valid. No error occurs. Output y matches input size.

3. Final Answer:

   A vector of size 44100x1 -> Option D

4. Quick Check:

   filter output length = input length [OK]

Hint: Filter output length equals input length [OK]

Common Mistakes:

• Assuming output length changes with filter order
• Expecting error from coefficients
• Confusing vector with scalar output

4. Identify the error in this MATLAB code that tries to add an echo effect to an audio signal x:

delay = 5000;
echoGain = 0.6;
y = x;
y(delay:end) = y(delay:end) + echoGain * x(1:end-delay);

medium

A. The length of y(delay:end) and x(1:end-delay) do not match.

B. The variable echoGain should be greater than 1.

C. The indexing delay:end is invalid for vector y.

D. The echo effect requires a filter function, not direct addition.

Solution

1. Step 1: Check vector lengths in addition

   The segment y(delay:end) has length length(y) - delay + 1. The segment x(1:end-delay) has length length(x) - delay. They differ by 1, causing a dimension mismatch error.

2. Step 2: Eliminate other options

   A: Indexing is valid. B: echoGain=0.6 <1 is appropriate for decaying echo. D: Direct indexing addition works for simple echo.

3. Final Answer:

   The length of y(delay:end) and x(1:end-delay) do not match. -> Option A

4. Quick Check:

   len(y(delay:end))=N-d+1 vs len(x(1:end-d))=N-d [OK]

Hint: Check lengths of vectors before adding [OK]

Common Mistakes:

• Assuming echoGain must be >1
• Thinking filter is required for echo
• Ignoring vector length mismatch

5. You want to create a MATLAB function that adds an echo effect to an audio signal x sampled at Fs. The echo delay is 0.3 seconds and echo gain is 0.5. Which code correctly implements this?

hard

A. function y = addEcho(x, Fs)
delaySamples = round(0.3 * Fs);
y = x;
y(delaySamples+1:end) = y(delaySamples+1:end) + 0.5*x(1:end-delaySamples);
end

B. function y = addEcho(x, Fs)
delaySamples = round(0.3 * Fs);
y = filter([1 zeros(1,delaySamples-1) 0.5], 1, x);
end

C. function y = addEcho(x, Fs)
delaySamples = round(0.3 * Fs);
y = x + 0.5 * [zeros(delaySamples,1); x];
end

D. function y = addEcho(x, Fs)
delaySamples = round(0.3 * Fs);
y = zeros(length(x)+delaySamples,1);
y(1:length(x)) = x;
y(delaySamples+1:end) = y(delaySamples+1:end) + 0.5*x;
end

Solution

1. Step 1: Understand echo effect implementation

   Echo is created by adding a delayed and scaled copy of the signal to itself. Using filter with coefficients [1 zeros(1,delaySamples-1) 0.5] adds original plus delayed echo.

2. Step 2: Evaluate each option

   function y = addEcho(x, Fs) delaySamples = round(0.3 * Fs); y = zeros(length(x)+delaySamples,1); y(1:length(x)) = x; y(delaySamples+1:end) = y(delaySamples+1:end) + 0.5*x; end extends length but adds full x without delay in echo part. function y = addEcho(x, Fs) delaySamples = round(0.3 * Fs); y = x; y(delaySamples+1:end) = y(delaySamples+1:end) + 0.5*x(1:end-delaySamples); end correctly indexes x to match lengths. function y = addEcho(x, Fs) delaySamples = round(0.3 * Fs); y = x + 0.5 * [zeros(delaySamples,1); x]; end adds vectors of different lengths causing error. function y = addEcho(x, Fs) delaySamples = round(0.3 * Fs); y = filter([1 zeros(1,delaySamples-1) 0.5], 1, x); end correctly uses filter with proper coefficients.

3. Final Answer:

   function y = addEcho(x, Fs) delaySamples = round(0.3 * Fs); y = filter([1 zeros(1,delaySamples-1) 0.5], 1, x); end -> Option B

4. Quick Check:

   Use filter with delayed coefficients for echo [OK]

Hint: Use filter with delayed coefficients for echo [OK]

Common Mistakes:

• Adding vectors of different lengths directly
• Forgetting to delay the echo signal
• Incorrect indexing causing errors