Why This Piece Exists
of the Fourier Rework — extra like an instinct piece primarily based on what I’ve discovered from it and its utility in sound frequency evaluation. The aim right here is to construct instinct for a way the Fourier Rework helps us get to frequency area options from time area options. We received’t get into heavy math and derivations; as a substitute, we’ll attempt to simplify the that means conveyed by the advanced equations.
Earlier than we get into the Fourier Rework, it’s best to have a fundamental understanding of how digital sound is saved — particularly sampling and quantization. Let me rapidly cowl it right here so we’re on the identical web page.
Sound in the true world is a steady wave — air stress altering easily over time. However computer systems can’t retailer steady issues. They want numbers, discrete values. To retailer sound digitally, we do two issues.
First, sampling — we take “snapshots” of the sound wave’s amplitude at common intervals. What number of snapshots per second? That’s the sampling price. CD-quality audio takes 44,100 snapshots per second (44.1 kHz). For speech in ML pipelines, 16,000 per second (16 kHz) is widespread and principally enough. I’ve labored with 16 kHz speech knowledge extensively, and it captures just about every little thing that issues for speech. The important thing thought is that we’re changing a clean steady wave right into a collection of discrete cut-off dates.
Second, quantization — every snapshot must document how loud the wave is at that second, and with how a lot precision. That is the bit depth. With 16-bit audio, every amplitude worth could be one among 65,536 doable ranges (2¹⁶). That’s greater than sufficient for the human ear to note any distinction from the unique. With solely 8-bit, you’d have simply 256 ranges — the audio would sound tough and grainy as a result of the hole between the true amplitude and the closest storable worth (this hole known as quantization error) turns into audible.
After sampling and quantization, what we now have is a sequence of numbers — amplitude values at evenly spaced time steps — saved within the laptop. That’s our time area sign. That’s g(t). And that’s what the Fourier Rework takes as enter.
I’ve spent a great period of time working hands-on with audio knowledge preprocessing and mannequin coaching, principally coping with speech knowledge. Whereas this piece builds every little thing from first rules, plenty of what’s written right here comes from truly operating into this stuff in actual pipelines, not simply textbook studying.
Additionally a promise — no AI slop right here. Let’s get into it.
The Setup: What We’re Beginning With
The unique audio sign — for advanced sounds (together with harmonic ones) just like the human voice or musical devices — is usually made up of a mixture of frequencies: constituent frequencies, or a superposition of frequencies.
The continual sound we’re speaking about is within the time area. It will be an amplitude vs. time graph. That’s how the sampled factors from the unique sound are saved in a pc in digital format.
The Fourier Rework (FT) is the mechanism by which we convert that graph from the time area (X-axis → Time, Y-axis → Amplitude) right into a frequency area illustration (X-axis → Frequency, Y-axis → Amplitude of contribution).
In case you’ve ever used librosa.stft() or np.fft.rfft() in your ML pipeline and puzzled what’s truly occurring below the hood while you go from uncooked audio to a spectrogram — that is it. The Fourier Rework is the muse beneath all of it.
Let’s discuss extra at an instinct degree about what we’re aiming for and the way the Fourier Rework delivers it. We’ll attempt to perceive this in an organized method.
Our Purpose
We need to discover the values of these frequencies whose mixture makes up the unique sound. By “authentic sound,” I imply the digital sign that we’ve saved by sampling and quantization by way of an ADC into our digital system. In easier phrases – we need to extract the constituent frequencies from which the advanced sound consists.
It’s analogous to having a bucket by which all colors are blended, and we need to segregate the constituent colors. The bucket blended with colors is the unique audio sign. The constituent colors are the constituent frequencies.
We wish a graph that simply tells us which frequencies have what amplitude of contribution in making the unique sound. The x-axis of that graph ought to have all of the frequency values, and the y-axis ought to have the amplitude of contribution corresponding to every frequency. The frequencies which are truly current within the sign will present up as peaks. The whole lot else will likely be close to zero.
Our enter can be the amplitude-time graph, and the output can be the amplitude-frequency graph from the Fourier Rework.
It’s apparent that since these graphs look so totally different, there can be arithmetic concerned. And to be sincere, superior mathematical instruments just like the Fourier Rework and complicated numbers are used to transform from our enter (time area graph) to our output (frequency area graph). However to get the instinct of why the Fourier Rework does the job accurately, it’s important to grasp what the Fourier Rework does such that our purpose is achieved. Then we’ll get to know how it helps us obtain it at an instinct degree.
The WHAT, the HOW, and the WHY.
The WHAT: What Does FT Really Do?
In answering the WHAT, we don’t have to see what math is happening inside — we simply need to know what enter it takes and what output it offers. We’ll deal with it like a black field.
Right here’s the factor: the enter to the FT is the complete authentic audio sign g(t), the whole time area waveform. We consider the FT at a selected frequency worth f, and the output for that frequency f is a single advanced quantity. This advanced quantity known as the Fourier coefficient for frequency f.
The subsequent query is: what is that advanced quantity that the FT outputs? What will we get from it?
From this advanced quantity, we extract two issues:
Magnitude = √(Real² + Imaginary²) — this tells us the amplitude of contribution of frequency f within the authentic sign. A excessive magnitude means f is strongly current within the authentic audio. A low magnitude means it’s barely there or not there in any respect.
Part = arctan(Imaginary / Actual) — this tells us the section offset of that frequency element. It signifies the place in its cycle that frequency begins. We’ll speak about section correctly later; don’t fear about it proper now. Simply know that this info additionally comes out of the identical advanced quantity.
What occurs is that we do that for each frequency we care about. For every f, we get one advanced quantity, extract the magnitude, and plot it. The gathering of all these (frequency, magnitude) pairs offers us the frequency area graph. That’s the WHAT.
Let’s see HOW that advanced quantity truly comes about — what’s the mechanism contained in the FT that produces it?
The HOW: How Does FT Compute This?
Right here’s the place issues get actually stunning, consider me.
The Winding Machine
The core thought is that we wrap the unique sign round a circle within the advanced airplane. The pace at which we wrap is dependent upon the enter frequency f.
Mathematically, for a given frequency f, we compute:
g(t) · e^(−2πift)
at each cut-off date t, and plot the outcome on the advanced airplane (actual axis, imaginary axis). Let’s break this down, as a result of it’s important to grasp how one can visualize and interpret what’s occurring right here.
Right here’s an essential factor to visualise: within the authentic g(t) graph, as time t will increase, we’re merely transferring from left to proper alongside the time axis — it’s a straight line, and we by no means come again. However within the advanced airplane, we’re transferring in a circle across the origin (0,0). As time progresses, we hold coming again to the identical angular positions — each time one full loop is accomplished, we begin over from the identical angle. The pace at which one full circle is accomplished is dependent upon f: one full rotation occurs when 2πtf = 2π, which suggests t·f = 1, so it takes 1/f seconds to finish one loop. Greater f → quicker looping. Decrease f → slower looping.
The time area graph is a one-way journey left to proper. The advanced airplane graph is a round journey that retains looping — and the speed of looping is managed by the enter frequency f.
You may suppose: since we hold coming again to the identical angular positions, does the second loop hint the very same path as the primary? Within the time area, every particular person constituent frequency is a repeating sine wave, proper? The 300 Hz element repeats each 1/300 seconds, the 700 Hz element repeats each 1/700 seconds. Every one individually has a clear repeating sample. After we wind g(t) across the advanced airplane, shouldn’t the trail from 0 to T (one interval, T = 1/f) and from T to 2T be precisely the identical? Shouldn’t the loops overlap completely?
No. And it is a delicate however essential factor to grasp early.
The person constituent frequencies inside g(t) do repeat — sure. However g(t) itself is just not a single frequency. It’s a superposition of a number of frequencies blended collectively. Though the angular place within the advanced airplane resets each 1/f seconds (the e^(−2πift) half completes one full loop), the gap from the origin — which is g(t) — is totally different at time t versus time t + 1/f. That’s as a result of g(t) has different frequency parts in it that don’t repeat on the identical price as f. The worth of g(t) on the identical angular place adjustments from one loop to the subsequent.
Every loop traces a barely totally different path within the advanced airplane. That is why, after we compute the Centre of Mass later, we compute it over the complete path for the complete period — not only one loop. If g(t) occurred to be a single pure sine wave at precisely frequency f and nothing else, then sure, each loop can be equivalent. However for any real-world sign with a number of frequencies, every loop is totally different, and we have to take into account all of them.
Maintain this in thoughts — it’ll make extra sense as soon as we get to the COM part beneath.
At any explicit time t:
g(t) is the amplitude of the unique sign at that second — this turns into the gap from the origin within the advanced airplane. Consider it because the magnitude of a fancy quantity.
e^(−2πift) offers the angle — particularly, an angle of (−2πtf) radians measured clockwise from the constructive actual axis.
At every time t, we’re putting some extent at distance g(t) from the origin, at an angle decided by 2πtf.
As time progresses, the angle retains rotating (as a result of t will increase), and the gap from the origin retains altering (as a result of g(t) adjustments with the audio sign). The result’s a path — a curve within the advanced airplane.
We are able to interpret this as wrapping or winding the unique sound sign g(t) round a circle, the place the pace of winding relies upon upon the enter frequency f. Greater f means the curve wraps round quicker. Decrease f means slower wrapping. One full circle is accomplished when t·f = 1, so the time interval of 1 full rotation is 1/f.
To visualise how this winding occurs at totally different frequencies, see this video — it can present the advanced graph form within the advanced airplane at totally different frequencies → 3Blue1Brown — However what’s the Fourier Rework? (https://www.youtube.com/watch?v=spUNpyF58BY). Probably the greatest assets on the market for constructing this instinct.
The Centre of Mass (COM)
Right here’s the place the magic occurs. As soon as we now have this wound-up curve within the advanced airplane, we calculate its Centre of Mass (COM).
Consider the wound-up curve as if it has uniform mass density, like a wire. The COM is the one level that represents the typical place of the complete curve. We wish the coordinates (Actual, Imaginary) of this COM. Let’s see how we truly calculate this.
Our authentic sound g(t), as a digitally saved sign in a pc, received’t be steady — we might have sampled factors of the unique sound. The corresponding sampled factors can be there on the advanced airplane too after making use of g(t)·e^(−2πift). The extra sampled factors there are within the authentic audio, the extra corresponding factors there can be on the advanced airplane.
A fast word earlier than the formulation: what we’ve been discussing up to now — the winding, the round movement, the COM — all of that’s the identical whether or not we’re speaking in regards to the steady model (with integrals) or the discrete model (with summations). The core idea of what the Fourier Rework does doesn’t change. Don’t get confused while you see a summation (Σ) in a single formulation and an integral (∫) in one other — they’re doing the identical factor conceptually. Summation is for our finite sampled factors; the integral is for the theoretical steady case. For constructing instinct, you’ll be able to consider both one — the thought is equivalent. Simply totally different instruments for a similar job.
For our discrete digital sign with N sampled factors, the COM coordinates are:
COM = (1/N) Σ g(t_n) · e^(-2πit_n·f)
That is the discrete model – and that is precisely what’s occurring while you name np.fft.rfft() or np.fft.fft() in Python. It’s computing this winding + COM calculation for all frequencies without delay. That one perform name is doing this whole course of throughout each frequency bin concurrently.
Now simply think about if this isn’t carried out digitally. In that case, we don’t want sampled factors and we will work on a steady perform. Which means we can have infinite steady factors of authentic audio and corresponding infinite factors on the advanced airplane. As a substitute of summation, we will combine:
ĝ(f) = ∫ g(t) · e^(-2πift) dt
Integration over limits → t₁ and t₂ (time period of authentic sound), integration over → g(t)·e^(-2πift), and the output is the advanced Fourier coefficient for that frequency f. That is the continual Fourier Rework formulation. In observe we all the time work with the discrete model since we’re coping with digital audio, however the steady type is sweet to know as a result of it exhibits the identical thought with out the distraction of indices and array lengths.
One factor price noting – the boundaries t₁ and t₂ matter. The ultimate COM you get truly is dependent upon how a lot of the sign you’re together with. A distinct time phase may give a distinct COM for a similar frequency. For this text, we’re making use of FT to the complete sign, so t₁ and t₂ are merely the beginning and finish of our whole audio. However while you later get into STFT (Quick-Time Fourier Rework), you’ll see that intentionally selecting brief time segments and making use of FT to every one is strictly the thought – and that’s the place window measurement turns into a design determination.
Now after we get the COM coordinates, we calculate its distance from the origin:
Magnitude = √(Real² + Imaginary²)
This magnitude is the amplitude of contribution of frequency f within the authentic audio sign. That’s what will get plotted because the y-value for this frequency within the frequency area graph.
The instinct for what this magnitude means: if the COM is at a big distance from the origin, that frequency has a robust contribution within the authentic sign. If the COM is sitting close to or across the origin, that frequency is barely current or not current in any respect. The gap from origin is instantly telling us how a lot that frequency issues.
And keep in mind what we mentioned earlier in regards to the loops not overlapping – that is the place it pays off. The COM averages over all these barely totally different loops, and that averaging is what makes the non-matching frequencies cancel out (their contributions level in several instructions throughout loops and sum to close zero) whereas the matching frequencies pile up (their contributions constantly level in the identical route throughout loops).
Why the COM Works: The Key Perception
That is the half that makes the entire thing click on. Learn this rigorously.
When the winding frequency f matches a constituent frequency of the sign, one thing particular occurs. The wound-up curve turns into lopsided — the factors pile up on one facet of the advanced airplane. The COM lands removed from the origin. Excessive magnitude. We detect that frequency.
When f does not match any constituent frequency, the wound-up curve distributes roughly evenly across the origin. Factors on one facet get cancelled out by factors on the other facet. The COM lands close to the origin. Low magnitude. That frequency isn’t actually current.
Match → lopsided → COM removed from origin → peak within the frequency area.
No match → balanced → COM close to origin → flat within the frequency area.
That’s it. That’s how the Fourier Rework figures out what frequencies are inside the unique sign.
Labored Instance: Strolling By means of the Numbers
Let’s make this concrete with precise numbers. That is the place the instinct turns into rock strong — belief me on this one.
Setup: Suppose our authentic audio sign is:
g(t) = sin(2π·300·t) + sin(2π·700·t)
This can be a sign made up of precisely two frequencies: 300 Hz and 700 Hz. In the true world, this may sound like two pure tones enjoying concurrently. We all know the reply already — the frequency area graph ought to present peaks at 300 and 700, and nothing else. Let’s see if the FT will get it proper.
We apply the Fourier Rework at three frequencies: f = 300 Hz, f = 700 Hz, and f = 500 Hz.
FT at f = 300 Hz (a constituent frequency)
We wind g(t) across the advanced airplane at 300 rotations per second.
Take into consideration what occurs — the 300 Hz element of g(t) is rotating at the very same pace as our winding. Due to this, the 300 Hz a part of the sign constantly lands on the identical facet of the advanced airplane. It doesn’t cancel itself out. The wound-up curve turns into closely lopsided in a single route.
What in regards to the 700 Hz element? It’s rotating at a distinct pace than our 300 Hz winding. Over time, it traces out a roughly symmetric path across the origin and averages out to close zero. It doesn’t contribute to the lopsidedness.
Outcome: The COM is much from the origin. The magnitude is excessive. The frequency area graph will get a tall peak at f = 300 Hz. Appropriate — 300 Hz is certainly a constituent frequency.
FT at f = 700 Hz (the opposite constituent frequency)
Identical logic, simply reversed. The 700 Hz element of g(t) matches the winding pace, so it piles up on one facet. The 300 Hz element, being at a distinct pace, averages out.
Outcome: The COM is much from the origin. Excessive magnitude. A tall peak at f = 700 Hz. Appropriate once more.
FT at f = 500 Hz (NOT a constituent frequency)
We wind g(t) at 500 rotations per second. Right here’s the factor — neither the 300 Hz element nor the 700 Hz element matches this winding pace. Each of them hint roughly symmetric paths across the origin within the advanced airplane. Nothing piles up constantly on one facet. The whole lot simply cancels out; the curve is just about centered across the origin.
Outcome: The COM may be very near the origin. The magnitude is close to zero. The frequency area graph is flat at f = 500 Hz — accurately telling us this frequency is just not current within the sign.
The Frequency Area Graph
After doing this for all frequencies, our frequency area graph would present precisely two sharp peaks — one at 300 Hz and one at 700 Hz — with every little thing else close to zero. We have now efficiently decomposed g(t) into its constituent frequencies. That’s the Fourier Rework doing its job.
The color bucket analogy holds completely: we had a combination (300 Hz + 700 Hz blended collectively within the time area), and the Fourier Rework segregated the constituent colors.
Seeing It in Code
For individuals who need to see this working in Python — right here’s the labored instance in precise code. It’s actually just a few traces:
import numpy as np
# Create the sign: 300 Hz + 700 Hz
sr = 8000 # sampling price
t = np.linspace(0, 1, sr, endpoint=False) # 1 second of audio
g = np.sin(2 * np.pi * 300 * t) + np.sin(2 * np.pi * 700 * t)
# Apply Fourier Rework - that is doing the winding + COM for all frequencies without delay
fft_result = np.fft.rfft(g)
# Get magnitudes (amplitude of contribution for every frequency)
magnitudes = np.abs(fft_result)
# Get the frequency values corresponding to every bin
freqs = np.fft.rfftfreq(len(g), d=1/sr)
# The peaks in magnitudes will likely be at 300 Hz and 700 Hz
# The whole lot else will likely be close to zeroThat’s it. np.fft.rfft(g) is doing the complete winding + COM course of we mentioned above – for each frequency bin concurrently. The np.abs() extracts the magnitude (distance of COM from origin), and the np.angle() would provide the section offset should you wanted it. The rfft particularly offers you solely the helpful half of the spectrum (as much as the Nyquist frequency) because the different half is a mirror – should you’ve learn the aliasing article, you recognize why.
Part: The Hidden Variable
Let’s speak about one thing that confused me for some time — the section. This idea is less complicated to understand if you have already got some understanding of section and section distinction by way of waves and sinusoidal indicators, however I’ll attempt to clarify what I understood.
I do know plenty of ML audio pipelines work with magnitude spectrograms solely and throw the section away totally. That’s positive for a lot of duties — however understanding what section is and what you’re discarding offers you a deeper understanding of the sign. And there are duties the place section issues (speech synthesis, audio reconstruction, vocoder design), so this part is price studying even should you’re solely doing magnitude-based function extraction proper now.
The COM we get from the FT is a fancy quantity. It has a magnitude (distance from the origin) and in addition an angle related to it:
Part = arctan(Imaginary(COM) / Actual(COM))
That angle tells us the section offset of the frequency element f because it exists inside the unique sign. In easy phrases, it tells you the place in its cycle that frequency element begins at t = 0.
A False impression I Had
I initially thought that for constituent frequencies, this section would all the time be 0. If a frequency is a part of the unique sign, the COM ought to simply lie on the true axis, proper? Part 0, most sync, all that. It is smart intuitively, no?
That’s not true, and right here’s why.
If the unique sign is g(t) = sin(2π·300·t + π/4), the frequency 300 Hz is totally a constituent frequency — it’s actually the one frequency within the sign. However its section offset is π/4, not 0. The 300 Hz element doesn’t begin at zero amplitude at t = 0; it begins shifted by π/4.
The FT will accurately output a excessive magnitude at f = 300 Hz, and the angle of the advanced quantity will likely be π/4, recovering the precise section with which the 300 Hz element exists within the sign.
Part is 0 provided that the element occurs to start out at precisely the correct reference level at t = 0. In any other case, it may be something. The magnitude tells you the way a lot of that frequency is current. The section tells you the place in its cycle it begins. Each items of knowledge come from the identical advanced quantity.
In code, you’d get these individually:
magnitude = np.abs(fft_result) # how a lot of every frequency
section = np.angle(fft_result) # the place in its cycle every frequency beginsWhenever you compute a magnitude spectrogram (which is what most ML pipelines do), you’re preserving the primary and discarding the second. Now at the least you recognize what you’re throwing away.
For Non-Constituent Frequencies
For frequencies that aren’t a part of the unique sign (like f = 500 Hz in our labored instance), the magnitude is close to zero. The section you get on this case is basically meaningless – it’s the angle of a near-zero vector pointing in some arbitrary route. Consider it as noise. The route doesn’t imply something when the vector has no size.
It’s fairly intuitive when you concentrate on it: for a non-constituent frequency, regardless of the COM coordinates come out to be, they’re so near the origin that the angle is simply numerical noise, not significant details about the sign.
Why FT Handles Part Routinely (This One Actually Confused Me)
Okay, so it is a delicate level that took me some time to get. And I need to clarify it clearly as a result of it’s the sort of factor that bugs you when you begin fascinated by it.
Right here’s the query: the FT solely takes frequency f as enter, proper? We don’t give it a section angle. However for a selected enter frequency, we may get totally different correlations if we differ the section alignment between our take a look at wave and the unique sign. So how does FT discover the “finest” section – the one that offers the utmost doable magnitude for enter frequency f?
The reply: FT doesn’t search or optimize over section in any respect. It doesn’t have to.
Right here’s why, and the secret is Euler’s formulation:
e^(-2πift) = cos(2πft) – i·sin(2πft)
After we compute FT at frequency f, we’re concurrently correlating the sign with each cos(2πft) and sin(2πft). The actual a part of the output captures the cosine correlation. The imaginary half captures the sine correlation.
Now right here’s the essential factor – any sinusoid at frequency f with any arbitrary section φ could be decomposed as:
A·cos(2πft + φ) = A·cos(φ)·cos(2πft) – A·sin(φ)·sin(2πft)
No matter what section the element has within the authentic sign, the FT mechanically captures it:
The actual half picks up A·cos(φ) — the cosine correlation. The imaginary half picks up A·sin(φ) — the sine correlation. Magnitude = √(real² + imag²) = A — the true amplitude, no matter φ. Angle = arctan(imag/actual) = φ — recovers the precise section.
It’s like measuring the size of a vector by projecting it onto each the x-axis and y-axis. Irrespective of which route the vector factors, you all the time recuperate its full size by √(x² + y²). The advanced exponential is testing all phases concurrently as a result of cosine and sine collectively cowl all doable section angles — they’re orthogonal to one another.
No optimization. No looking. No iterating over section values. Simply the truth that cosine and sine are orthogonal and collectively they seize any section. The mathematics does it in a single shot.
That is the place I lastly understood why advanced numbers are used right here and never simply common correlation with a single sine wave. Euler’s formulation is doing one thing very intelligent — it’s correlating with two issues without delay, and the advanced quantity neatly packages each outcomes collectively.
Placing It All Collectively
Right here is the complete image of how we get from the time area to the frequency area:
1. Take the unique audio sign g(t) — our time area knowledge
2. Decide a frequency f
3. Wind g(t) across the advanced airplane at pace f utilizing g(t)·e^(−2πift)
4. Calculate the COM of the wound-up curve
5. The gap of the COM from the origin → amplitude of contribution of f
6. The angle of the COM → section offset of f
7. Plot the purpose (f, magnitude) on the frequency area graph
8. Repeat for all frequencies
The frequencies which are truly current within the authentic sign produce lopsided winding → COM removed from the origin → peaks within the graph. Frequencies that aren’t current produce balanced winding → COM close to the origin → flat areas.
After doing this throughout all frequencies, we now have the whole frequency area graph. The peaks inform us the constituent frequencies of the unique sound. That’s the Fourier Rework — decomposing a fancy sign into its constructing blocks.
The mathematics is a device to justify the instinct — the true understanding is within the winding, the Centre of Mass, and the best way the advanced exponential handles section mechanically by Euler’s formulation. As soon as these three issues click on, you get the Fourier Rework at an instinct degree, and the heavy math derivations are simply formalizing what you already perceive. And as soon as this clicks, you’ll see the FT all over the place in sign processing, and it’ll all begin making sense.
The WHY
Why does the Fourier Rework work? The intuitive reply is what we’ve constructed by this whole piece – matching frequencies create lopsided windings, non-matching frequencies create balanced ones that cancel out. The winding machine is basically a correlation detector – it measures how a lot the unique sign correlates with a pure sinusoid at every frequency. Excessive correlation means COM removed from origin which provides a peak, low correlation means COM close to origin and we get a flat area within the graph.
At its core, why this works rigorously would require heavy math derivation involving orthogonality of sinusoidal features and properties of advanced exponentials – which isn’t the aim of this piece. However the instinct we’ve constructed must be greater than sufficient to grasp what’s occurring and why the output is smart. It really works!
What Comes Subsequent
This piece covers the continual/conceptual Fourier Rework — the muse. In observe, while you work with digital audio in ML pipelines, you’re utilizing the DFT (Discrete Fourier Rework) and its quick implementation, the FFT. And while you compute spectrograms, you’re utilizing the STFT (Quick-Time Fourier Rework), which applies the FT to small overlapping home windows of the sign — that’s the place window measurement N, hop size, and overlap are available. However that’s a subject for an additional writeup.
All of that builds instantly on high of what we lined right here. The winding machine, the COM, the magnitude and section — it’s the identical mechanism, simply utilized to brief chunks of audio as a substitute of the entire thing without delay. If this piece clicked for you, the remainder will observe naturally. I’d write in regards to the DFT and STFT intimately later.
Thanks for the endurance should you’ve learn this far, and due to Grammarly for serving to with the enhancing.
Be at liberty to achieve out with any questions:
E mail: [email protected]
Twitter: @r4plh
GitHub: github.com/r4plh
LinkedIn: linkedin.com/in/r4plh
