The library provides functions to implement a very fast FFT specifically on the ESP32-S3 MCU with PSRAM. These functions are similar to the popular arduinoFFT but are many times faster due to the ESP32-S3 DSP support.
Hardware support: ESP32-S3 MCU with PSRAM.
-
Copy the two files to your project source folder and add #include esp32s3_fft.h to your header file.
-
In your main code instanciate the library: ESP32S3_FFT fft;
-
To prepare for an FFT operation, call fft.init(fft_size, fft_samples, spectral_output). * fft_size * is typically 512 but can be any value in powers of 2; Example: 128, 256, 512, 1024... fft_samples are the number of time domain samples to be converted to frequency domain. Input data samples can be of arbitrary length but FFT accuracy improves if the number of samples is a even multiple of fft_size. For info on spectral_output options, see below. The fft.init() function returns a pointer to a fft_table_t structure containing parameters for allocating input & output buffers.
-
Call fft.compute(float *source_bufr, float *output_bufr); to convert source data (audio or ???). The results of the FFT are written to output_bufr.
- num_original_samples: This is the same as the fft_samples parameter of the init() function.
- size_input_bufr: Allocate this number of SAMPLES for the data input buffer.
- num_sliding_frames: Allocate the output buffer in PSRAM using the formula: num_sliding_frames * fft_size * sizeof(float).
- hop_size: The size of the sliding window (normally 50% of the fft_size.
FFT data can be processed and returned in three different ways:
-
SPECTRAL AVERAGE - FFT data is returned to the output buffer whose legnth is equal to 'fft_size'. The resulting data is the average of all transforms across the input data using a 50% sliding window.
-
SPECTRAL NO SLIDING - FFT data is returned in multiple blocks without sliding window overlap.
-
SPECTRAL SLIDING - FFT data is returned using a sliding window with a 50% overlap.
Google "FFT sliding window" for an explaination of the technique.