FFT size is the number of time-domain samples used in a Fast Fourier Transform calculation. It directly controls how finely you can distinguish between frequencies in the result. Common FFT sizes are powers of two (256, 512, 1024, 2048, 4096, 8192, 16384, 65536), and choosing the right one is a balancing act between frequency detail and time precision.
How FFT Size Controls Frequency Resolution
The core relationship is simple. The width of each frequency “bin” in your FFT result equals the sample rate divided by the FFT size:
Bin width = Sample rate / FFT size
So if you’re working at a 48 kHz sample rate with an FFT size of 16,384 samples, each bin covers 48,000 / 16,384 = about 2.93 Hz. That means you can distinguish between two tones only if they’re at least 2.93 Hz apart. Double the FFT size to 32,768 and you cut that bin width in half, giving you roughly 1.46 Hz resolution.
There’s another way to think about this. The bin width is also equal to 1 divided by the duration of your signal window in seconds. If you capture 16,384 samples at 48 kHz, that’s about 0.34 seconds of audio, giving you a frequency resolution of roughly 1 / 0.34 = 2.93 Hz. The math works out the same either way: longer recordings give you finer frequency detail.
The total number of usable frequency bins in the output is FFT size / 2 + 1, spanning from 0 Hz (DC) up to half the sample rate. At 48 kHz with a 16,384-point FFT, that’s 8,193 bins covering DC to 24 kHz.
The Time vs. Frequency Trade-Off
Here’s the catch: a larger FFT size means you need a longer chunk of signal to analyze, which blurs your ability to pinpoint when something happened in time. If you’re analyzing a drum hit followed by a cymbal crash, a massive FFT window might smear those two events together into one blended spectrum. You’d get beautifully detailed frequency information, but you’d lose the ability to see each hit as a separate event.
Shorter FFT sizes do the opposite. They give you sharp time resolution so you can track rapid changes in the signal, but the frequency bins become wider and coarser. Each doubling of FFT size doubles the frequency resolution and extends the lowest resolvable frequency down by one octave, at the cost of halving your time precision. This inverse relationship is fundamental and unavoidable. It’s not a flaw in the algorithm; it’s a mathematical property of how signals work.
Why Powers of Two Matter
The “fast” in Fast Fourier Transform refers to a computational shortcut that works when the number of samples is a power of two (512, 1024, 2048, etc.). The algorithm splits the problem into smaller and smaller halves, dramatically reducing the number of calculations needed. Some software requires a power-of-two size. Others can handle arbitrary lengths but run faster with powers of two. In practice, nearly everyone sticks to these standard sizes.
Choosing an FFT Size in Practice
The right FFT size depends entirely on what you’re trying to see in the signal. For audio work at a 48 kHz sample rate, here’s what different sizes get you:
- 1,024 samples: Bin width of about 47 Hz. Good for fast, responsive spectrum displays, but far too coarse to distinguish individual notes in the bass range.
- 4,096 samples: Bin width of about 11.7 Hz. A common choice for real-time audio visualizers where you want a reasonable balance of speed and detail.
- 16,384 samples: Bin width of about 2.93 Hz. Enough resolution to separate individual musical notes across most of the frequency range.
- 65,536 samples: Bin width under 1 Hz. Used for precise acoustic measurements and detailed analysis where time response doesn’t matter much.
For sub-bass frequencies (below about 40 Hz), you need large FFT sizes because small windows simply can’t capture enough cycles of a low-frequency wave to identify it. A 20 Hz tone completes one full cycle every 50 milliseconds, so your analysis window needs to be at least that long to even “see” it. At 48 kHz, that’s a minimum of 2,400 samples, pushing you toward an FFT size of 4,096 or larger.
Audio Precision notes that an FFT length of 65,536 samples at a sample rate of 65,536 Hz produces a bin width of exactly 1 Hz, which is a convenient configuration for precision testing equipment.
Zero Padding: More Bins, Not More Resolution
A common technique is to pad your signal with zeros to increase the FFT size without collecting more data. If you have 1,024 real samples and pad them out to 4,096, the FFT output will have four times as many bins and the spectrum will look smoother. But this is interpolation, not genuine resolution improvement. The additional bins fill in the gaps between the original frequency points, making peaks easier to spot visually, but they don’t let you separate two closely spaced frequencies that the original 1,024 samples couldn’t distinguish.
To actually increase frequency resolution, you need to record a longer signal. More real data means more information about the signal’s frequency content. Zero padding just makes the existing information look prettier on a graph.
Spectral Leakage and Window Functions
FFT math assumes that your block of samples repeats infinitely. When the signal doesn’t complete exact whole cycles within your FFT window (which is almost always the case), the result shows energy smeared across neighboring bins. This artifact is called spectral leakage, and it makes peaks look broader and messier than they should be.
Window functions fix this by tapering the signal to zero at both edges of the block before the FFT runs. The most common is the Hanning window, which smoothly fades the signal at each end. Applying it dramatically reduces leakage, producing sharp, clean peaks instead of broad smeared ones. The trade-off is that windowing slightly widens the main peak and reduces the effective resolution compared to a perfectly periodic signal.
Different windows serve different purposes. A flat-top window gives the most accurate amplitude readings but poor frequency resolution, making it ideal when you need to measure exactly how loud a known tone is. A Hanning window offers a good general-purpose balance. The choice of window interacts with your FFT size: a larger FFT compensates for some of the resolution lost to windowing, which is another reason analysts often reach for bigger sizes when precision matters.
FFT Size Beyond Audio
The same principles apply in any domain that uses FFTs. In wireless communications, FFT size determines how many subcarrier frequencies can be packed into a channel. In vibration analysis for machinery, larger FFT sizes help engineers pinpoint the exact rotational frequency causing a problem. In medical imaging, FFT size affects how finely spatial frequencies are resolved. The core trade-off never changes: more frequency detail costs you time detail, and the bin width always equals the sample rate divided by the number of points.