|
| 1 | +#include "mfcc.h" |
| 2 | + |
| 3 | +MFCC::MFCC(int num_mfcc_coeffs, int frame_size, int num_fft_points) : |
| 4 | + _num_mfcc_coeffs(num_mfcc_coeffs), |
| 5 | + _frame_size(frame_size), |
| 6 | + _num_fft_points(num_fft_points), |
| 7 | + _hanning_window(NULL), |
| 8 | + _mel_filterbank(NULL) |
| 9 | +{ |
| 10 | +} |
| 11 | + |
| 12 | +MFCC::~MFCC() |
| 13 | +{ |
| 14 | + if (_hanning_window != NULL) { |
| 15 | + delete [] _hanning_window; |
| 16 | + _hanning_window = NULL; |
| 17 | + } |
| 18 | + if (_mel_filterbank != NULL) { |
| 19 | + delete [] _mel_filterbank; |
| 20 | + _mel_filterbank = NULL; |
| 21 | + } |
| 22 | +} |
| 23 | + |
| 24 | +int MFCC::init() |
| 25 | +{ |
| 26 | + // Initialize the hanning window and Mel filterbank, similar to DSPPipeline::init |
| 27 | + |
| 28 | + return 1; |
| 29 | +} |
| 30 | + |
| 31 | +void MFCC::extract_mfcc(const int16_t* input, float32_t* output) |
| 32 | +{ |
| 33 | + int16_t windowed_input[_frame_size]; |
| 34 | + int16_t fft_q15[_frame_size * 2]; |
| 35 | + |
| 36 | + // Apply the MFCC pipeline: Hanning Window + FFT |
| 37 | + arm_mult_q15(_hanning_window, input, windowed_input, _frame_size); |
| 38 | + arm_rfft_q15(&_S_q15, windowed_input, fft_q15); |
| 39 | + |
| 40 | + // Calculate the magnitude spectrum (similar to DSPPipeline::calculate_spectrum) |
| 41 | + // Compute the power spectrum |
| 42 | + arm_cmplx_mag_q15(fft_q15, fft_mag_q15, _frame_size / 2 + 1); |
| 43 | + |
| 44 | + // Apply Mel filterbank to the power spectrum (specific to MFCC) |
| 45 | + float32_t mel_energies[_num_mfcc_coeffs]; |
| 46 | + apply_mel_filterbank(fft_mag_q15, mel_energies); |
| 47 | + |
| 48 | + // Compute the logarithm of the mel energies |
| 49 | + for (int i = 0; i < _num_mfcc_coeffs; i++) { |
| 50 | + mel_energies[i] = logf(mel_energies[i]); |
| 51 | + } |
| 52 | + |
| 53 | + // Apply DCT (Discrete Cosine Transform) to obtain MFCC coefficients |
| 54 | + compute_mfcc(output, mel_energies); |
| 55 | +} |
| 56 | + |
| 57 | +void MFCC::apply_mel_filterbank(const int16_t* spectrum, float32_t* mel_energies) |
| 58 | +{ |
| 59 | + // Define the filter bank parameters |
| 60 | + int num_filter_banks = 13; |
| 61 | + int filter_bank_size = _frame_size / 2 + 1; // Half of the FFT size |
| 62 | + |
| 63 | + // Initialize the mel filterbank |
| 64 | + if (_mel_filterbank == NULL) { |
| 65 | + _mel_filterbank = new float32_t[num_filter_banks * filter_bank_size]; |
| 66 | + |
| 67 | + // Initialize the mel filterbank with appropriate filter shapes |
| 68 | + // You can use equations like Triangular, Hanning, or other shapes for filters |
| 69 | + // Fill _mel_filterbank with filter coefficients based on filter bank parameters |
| 70 | + // Ensure that the coefficients sum to 1 for each filter |
| 71 | + // This step is essential and depends on your specific filterbank design. |
| 72 | + } |
| 73 | + |
| 74 | + // Apply the mel filter bank to the spectrum |
| 75 | + for (int i = 0; i < num_filter_banks; i++) { |
| 76 | + mel_energies[i] = 0.0; |
| 77 | + for (int j = 0; j < filter_bank_size; j++) { |
| 78 | + mel_energies[i] += _mel_filterbank[i * filter_bank_size + j] * spectrum[j]; |
| 79 | + } |
| 80 | + } |
| 81 | +} |
| 82 | + |
| 83 | +void MFCC::compute_mfcc(float32_t* mfcc_output, const float32_t* mel_energies) |
| 84 | +{ |
| 85 | + // Define the number of MFCC coefficients |
| 86 | + int num_mfcc_coeffs = 9; |
| 87 | + |
| 88 | + // Initialize the DCT matrix (you can precompute it) |
| 89 | + // It's a matrix of size num_mfcc_coeffs x 13 (number of filter banks) |
| 90 | + // You can find precomputed DCT matrices in DSP libraries or compute it manually |
| 91 | + |
| 92 | + // Compute the DCT of the mel energies to obtain MFCC coefficients |
| 93 | + for (int i = 0; i < num_mfcc_coeffs; i++) { |
| 94 | + mfcc_output[i] = 0.0; |
| 95 | + for (int j = 0; j < 13; j++) { |
| 96 | + mfcc_output[i] += mel_energies[j] * dct_matrix[i * 13 + j]; |
| 97 | + } |
| 98 | + } |
| 99 | +} |
| 100 | + |
0 commit comments