Audio Filter Design Fundamentals

From Analysis to Synthesis: The Role of Digital Filters

A digital filter is an algorithm that takes a sequence of numbers (our audio samples) and produces a new sequence where the frequency content has been altered. This is the fundamental building block for a vast range of audio effects:

• Equalizers (EQ): The bass and treble controls on a stereo are simple filters.
• Noise Reduction: Filters can be used to remove unwanted hiss or hum.
• Special Effects: Wah-wah pedals, phasers, and other effects are all based on dynamically changing filters.
• Crossovers in speakers: Filters are used to route low frequencies to the woofer and high frequencies to the tweeter.

In digital signal processing, there are two primary families of filters: Finite Impulse Response (FIR) and Infinite Impulse Response (IIR). Your first decision as a filter designer is to choose which family is right for your application.

1. The Two Families: FIR vs. IIR Filters

The names "Finite" and "Infinite" Impulse Response refer to the length of the filter's response to a single, brief input spike (an impulse). This mathematical property leads to very different practical trade-offs.

Let's summarize the key trade-offs in the context of our audio project:

The Bottom Line: If pristine audio quality and zero phase distortion are paramount, and you can afford the computational cost, choose FIR. If computational efficiency is the top priority and some phase distortion is acceptable, choose IIR.

2. Designing FIR Filters: The Window Method

The window method is an intuitive and popular technique for designing FIR filters. The process starts with an "ideal" filter and makes it practical.

The Concept

1. Start with the Ideal: Imagine a perfect "brick-wall" lowpass filter. Its frequency response is 1 in the passband and 0 in the stopband.
2. Find its Impulse Response: The impulse response of this ideal filter is the inverse Fourier transform of its frequency response. For a lowpass filter, this turns out to be the sinc function: $\sin (\pi x)/(\pi x)$.
3. The Problem: The sinc function is infinite in duration and non-causal (it starts before time zero). We cannot implement an infinitely long filter.
4. The Solution: We make it finite by "truncating" it—cutting it off after a certain number of samples. This is equivalent to multiplying the infinite impulse response by a finite-length "window" function.

This truncation introduces artifacts, most notably ripples in the passband and stopband (the Gibbs phenomenon). The shape of the window we use to truncate the sinc function allows us to control the trade-off between the sharpness of the filter's transition from passband to stopband and the amount of ripple.

FIR Design in a Nutshell

Given a set of filter specifications, the design process is:

1. Define Specifications:

   • Passband Edge Frequency ($f_p$): The end of the frequencies you want to keep.
   • Stopband Edge Frequency ($f_s$): The start of the frequencies you want to block.
   • Stopband Attenuation ($A_s$): How much you want to block the stopband frequencies, in dB.
   • Sampling Frequency ($f_{samp}$)

2. Choose a Window: Your required stopband attenuation ($A_s$) dictates the window type. For example:

   • Rectangular: ~21 dB (rarely used)
   • Hanning: ~44 dB
   • Hamming: ~53 dB
   • Blackman: ~74 dB

3. Determine Filter Length (Order): The required transition width ($\Delta f=f_s-f_p$) determines the filter length, $M$. The narrower the transition, the longer the filter needs to be. Each window has an approximate formula:

   • Hanning: $M\approx 3.1/(\Delta f/f_{samp})$
   • Hamming: $M\approx 3.3/(\Delta f/f_{samp})$
   • Blackman: $M\approx 5.5/(\Delta f/f_{samp})$

4. Calculate Coefficients: The filter coefficients, $h[n]$, are calculated by multiplying the ideal impulse response (a sinc function centered at the cutoff frequency) with the chosen window function.

3. Designing IIR Filters: The Analog Prototype Method

IIR filter design leverages the rich history of analog electronics. The strategy is to design a high-quality analog filter and then mathematically transform it into a digital equivalent. Your EE background in Laplace transforms and continuous-time systems will be very helpful here.

The Concept

1. Start in the Analog World: Choose a classic analog filter prototype that meets your desired frequency response shape.

*Four common analog filter prototypes. **Butterworth** is maximally flat in the passband. **Chebyshev** has a steeper roll-off but with ripple in the passband (Type I) or stopband (Type II). **Elliptic** is the steepest but has ripple in both bands. For audio, Butterworth is often a good starting point.*

2. Get the Analog Transfer Function, $H(s)$: Based on your specs (cutoff frequency, order) and chosen prototype, you determine the analog filter's transfer function in the Laplace domain (the s-plane).

3. Transform from s-plane to z-plane: Use a mathematical mapping to convert the analog transfer function $H(s)$ into a digital transfer function $H(z)$. The most common and robust method is the Bilinear Transform.

IIR Design with the Bilinear Transform in a Nutshell

1. Define Digital Specifications:

   • Passband edge ($f_p$), stopband edge ($f_s$), stopband attenuation ($A_s$), passband ripple ($A_p$).
   • Sampling frequency ($f_{samp}$).

2. Pre-warp Frequencies: The Bilinear Transform non-linearly "warps" the frequency axis. To get the desired digital cutoff frequency, you must aim for a different, "pre-warped" analog frequency. Convert your digital edge frequencies (${\omega }_p,{\omega }_s$) to pre-warped analog frequencies (${\Omega }_p,{\Omega }_s$) using the formula:
   $\Omega =\frac{2}{t_s}\tan \left(\frac{\omega }{2}\right)$
   where $\omega =2\pi f/f_{samp}$ and $t_s=1/f_{samp}$.

3. Design the Analog Filter: Choose a prototype (e.g., Butterworth). Using your pre-warped analog specs (${\Omega }_p,{\Omega }_s,A_p,A_s$), calculate the required filter order $n$ and find the analog transfer function $H(s)$.

4. Apply the Bilinear Transform: Substitute $s=\frac{2}{t_s}\frac{1-z^{-1}}{1+z^{-1}}$ into your $H(s)$. After a good deal of algebra, this will yield the digital transfer function $H(z)$.

5. Extract Coefficients: The numerator and denominator polynomials of $H(z)$ directly give you the b (feedforward) and a (feedback) coefficients for the filter's difference equation.

4. A Practical Design Example

Let's design a filter to a common audio specification.

Specification:

• Type: Lowpass filter
• Application: Isolate bass frequencies in an audio track.
• Sampling Rate: $f_{samp}=44.1$ kHz
• Cutoff Frequency: $f_p=250$ Hz
• Stopband Start: $f_s=500$ Hz
• Stopband Attenuation: $A_s\ge 40$ dB

Option 1: FIR Design (Window Method)

1. Window Choice: 40 dB attenuation is needed. A Hanning window (~44 dB) is a safe choice.
2. Filter Length: The normalized transition width is $\Delta f=(f_s-f_p)/f_{samp}=(500-250)/44100\approx 0.00567$.
   Using the Hanning formula: $M\approx 3.1/\Delta f\approx 3.1/0.00567\approx 547$. This is a 547-tap FIR filter.
3. Coefficients: We would generate the coefficients using software, but we know the process: create a sinc function for a cutoff halfway in the transition band (~375 Hz) and multiply it by a 547-point Hanning window.

Option 2: IIR Design (Bilinear Transform, Butterworth Prototype)

1. Pre-warp Frequencies:
   • ${\omega }_p=2\pi (250/44100)\approx 0.0356$ rad/sample $⟹{\Omega }_p=\frac{2}{t_s}\tan ({\omega }_p/2)\approx 1568$ rad/s.
   • ${\omega }_s=2\pi (500/44100)\approx 0.0712$ rad/sample $⟹{\Omega }_s=\frac{2}{t_s}\tan ({\omega }_s/2)\approx 3137$ rad/s.
2. Determine Order: For a Butterworth filter, the order $n$ is given by:
   $n\ge \frac{{\log }_{10}(10^{A_s/10}-1)}{2{\log }_{10}({\Omega }_s/{\Omega }_p)}$
   Plugging in our values ($A_s=40,{\Omega }_s=3137,{\Omega }_p=1568$):
   $n\ge \frac{{\log }_{10}(10^4-1)}{2{\log }_{10}(3137/1568)}\approx \frac{4}{2\times 0.301}\approx 6.64$
   We must round up to the next integer, so we need a 7th-order IIR filter.
3. Get H(z): We would find the transfer function $H(s)$ for a 7th-order analog Butterworth filter with cutoff ${\Omega }_p$, then apply the bilinear transform to get $H(z)$ and its coefficients.

The comparison is stark: a 547-tap FIR vs. a 7th-order IIR. The IIR filter will be vastly more computationally efficient, but will introduce some phase distortion. This perfectly illustrates the fundamental design trade-off.

Conclusion

You have now surveyed the entire landscape of digital filter design, from the high-level choice between FIR and IIR to the detailed mathematical steps required for two primary design methods. You are equipped to take a set of audio requirements and translate them into a concrete filter design.

Key Takeaways:

• FIR vs. IIR is the first major design choice, trading off phase linearity and guaranteed stability (FIR) against computational efficiency (IIR).
• The Window Method is a common way to design FIR filters by truncating the ideal impulse response. The choice of window controls the trade-off between transition width and stopband attenuation.
• The Bilinear Transform is a standard method for designing IIR filters by converting an analog prototype. It avoids aliasing but introduces frequency warping, which must be corrected by pre-warping the design specifications.
• For a given magnitude specification, IIR filters are almost always of a much lower order (more efficient) than their FIR counterparts.

Preview of the Next Lesson:

Designing a filter is the first half of the battle. The next step is to make it run. In our next lesson, you will learn to implement the designed filter on a real-time audio stream using its difference equation. We will take the coefficients we designed today and use them to process audio in our C++ application, finally hearing the results of our work.