viewof defbAudioRate = Inputs.select([8000, 16000, 48000], {value: 16000, label: "Sample rate (samples/s)"})
viewof defbAudioFft = Inputs.select([128, 256, 400, 512, 1024, 2048], {value: 400, label: "FFT length (samples)"})
viewof defbAudioOverlap = Inputs.select([0, 0.25, 0.5, 0.75, 0.875], {value: 0.5, format: value => `${Math.round(value * 100)}%`, label: "Frame overlap"})
viewof defbAudioWindow = Inputs.select(["Rectangular", "Hann", "Hamming"], {value: "Hann", label: "Window"})
viewof defbAudioTone = Inputs.range([200, 3000], {value: 997, step: 1, label: "Test tone (Hz)"})FFT And MFCC Parameter Explorer
FFT And MFCC Parameter Explorer
animation
audio
analytics-ml
Try Change FFT length (samples) from 400 to 1024.
Observe Compare Frame duration, Bin spacing, and Hop and analysis rate at 1024 samples.
Explain Explain why 1024 samples separate closer steady tones but blur short events.
Technical boundaries. The final row is a small leakage probe evaluating the selected tone at the nearest bin and its two neighbors. It is not a specification for a window.
FFT And MFCC Parameter Explorer
An FFT size is an evidence choice, not a quality slider. At sample rate , an -sample frame spans seconds and its bin spacing is . Longer frames separate closer steady tones but blur short events. Overlap reduces the hop between decisions and increases compute; it does not create new independent samples. A window reduces edge discontinuities at the cost of widening the response around a tone.
The final row is a small leakage probe, not a specification for a window: it evaluates the selected tone at the nearest bin and its two neighbors. Sweep the tone between bin centers and compare windows. Then choose the shortest frame that still separates the task’s relevant frequencies, measure compute and latency on the target, and keep the choice with the model record.
MFCCs add another bounded sequence: optional pre-emphasis, windowed spectrum, Mel-spaced filterbank energies, logarithm, and a discrete cosine transform. Record sample rate, frame, hop, window, filter count, coefficient count, normalization, and implementation version. Speech defaults are only a starting hypothesis for machinery, wildlife, or occupancy audio; validate that the compressed representation retains the events the IoT decision actually needs.