介绍：https://github.com/majianjia/nnom/blob/master/examples/rnn-denoise/README_CN.md
默认提供了一个wav的例子
#include <stdint.h>
#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <string.h>#include "nnom.h"
#include "denoise_weights.h"#include "mfcc.h"
#include "wav.h"// the bandpass filter coefficiences
#include "equalizer_coeff.h" #define NUM_FEATURES NUM_FILTER#define _MAX(x, y) (((x) > (y)) ? (x) : (y))
#define _MIN(x, y) (((x) < (y)) ? (x) : (y))#define NUM_CHANNELS 1
#define SAMPLE_RATE 16000
#define AUDIO_FRAME_LEN 512// audio buffer for input
float audio_buffer[AUDIO_FRAME_LEN] = {0};
int16_t audio_buffer_16bit[AUDIO_FRAME_LEN] = {0};// buffer for output
int16_t audio_buffer_filtered[AUDIO_FRAME_LEN/2] = { 0 };// mfcc features and their derivatives
float mfcc_feature[NUM_FEATURES] = { 0 };
float mfcc_feature_prev[NUM_FEATURES] = { 0 };
float mfcc_feature_diff[NUM_FEATURES] = { 0 };
float mfcc_feature_diff_prev[NUM_FEATURES] = { 0 };
float mfcc_feature_diff1[NUM_FEATURES] = { 0 };
// features for NN
float nn_features[64] = {0};
int8_t nn_features_q7[64] = {0};// NN results, which is the gains for each frequency band
float band_gains[NUM_FILTER] = {0};
float band_gains_prev[NUM_FILTER] = {0};// 0db gains coefficient
float coeff_b[NUM_FILTER][NUM_COEFF_PAIR] = FILTER_COEFF_B;
float coeff_a[NUM_FILTER][NUM_COEFF_PAIR] = FILTER_COEFF_A;
// dynamic gains coefficient
float b_[NUM_FILTER][NUM_COEFF_PAIR] = {0};// update the history
void y_h_update(float *y_h, uint32_t len)
{for (uint32_t i = len-1; i >0 ;i--)y_h[i] = y_h[i-1];
}// equalizer by multiple n order iir band pass filter.
// y[i] = b[0] * x[i] + b[1] * x[i - 1] + b[2] * x[i - 2] - a[1] * y[i - 1] - a[2] * y[i - 2]...
void equalizer(float* x, float* y, uint32_t signal_len, float *b, float *a, uint32_t num_band, uint32_t num_order)
{// the y history for each bandstatic float y_h[NUM_FILTER][NUM_COEFF_PAIR] = { 0 };static float x_h[NUM_COEFF_PAIR * 2] = { 0 };uint32_t num_coeff = num_order * 2 + 1;// i <= num_coeff (where historical x is involved in the first few points)// combine state and new data to get a continual x input. memcpy(x_h + num_coeff, x, num_coeff * sizeof(float));for (uint32_t i = 0; i < num_coeff; i++){y[i] = 0;for (uint32_t n = 0; n < num_band; n++){y_h_update(y_h[n], num_coeff);y_h[n][0] = b[n * num_coeff] * x_h[i+ num_coeff];for (uint32_t c = 1; c < num_coeff; c++)y_h[n][0] += b[n * num_coeff + c] * x_h[num_coeff + i - c] - a[n * num_coeff + c] * y_h[n][c];y[i] += y_h[n][0];}}// store the x for the state of next roundmemcpy(x_h, &x[signal_len - num_coeff], num_coeff * sizeof(float));// i > num_coeff; the rest data not involed the x historyfor (uint32_t i = num_coeff; i < signal_len; i++){y[i] = 0;for (uint32_t n = 0; n < num_band; n++){y_h_update(y_h[n], num_coeff);y_h[n][0] = b[n * num_coeff] * x[i];for (uint32_t c = 1; c < num_coeff; c++)y_h[n][0] += b[n * num_coeff + c] * x[i - c] - a[n * num_coeff + c] * y_h[n][c];y[i] += y_h[n][0];} }
}// set dynamic gains. Multiple gains x b_coeff
void set_gains(float *b_in, float *b_out, float* gains, uint32_t num_band, uint32_t num_order)
{uint32_t num_coeff = num_order * 2 + 1;for (uint32_t i = 0; i < num_band; i++)for (uint32_t c = 0; c < num_coeff; c++)b_out[num_coeff *i + c] = b_in[num_coeff * i + c] * gains[i]; // only need to set b.
}void quantize_data(float*din, int8_t *dout, uint32_t size, uint32_t int_bit)
{float limit = (1 << int_bit); for(uint32_t i=0; i<size; i++)dout[i] = (int8_t)(_MAX(_MIN(din[i], limit), -limit) / limit * 127);
}void log_values(float* value, uint32_t size, FILE* f)
{char line[16];for (uint32_t i = 0; i < size; i++) {snprintf(line, 16, "%f,", value[i]);fwrite(line, strlen(line), 1, f);}fwrite("\n", 2, 1, f);
}int main(int argc, char* argv[])
{wav_header_t wav_header; size_t size;char* input_file = "sample.wav";char* output_file = "filtered_sample.wav";FILE* src_file;FILE* des_file;char* log_file = "log.csv";FILE* flog = fopen(log_file, "wb");// if user has specify input and output files. if (argc > 1)input_file = argv[1];if (argc > 2)output_file = argv[2];src_file = fopen(input_file, "rb");des_file = fopen(output_file, "wb");if (src_file == NULL) {printf("Cannot open wav files, default input:'%s'\n", input_file);printf("Or use command to specify input file: xxx.exe [input.wav] [output.wav]\n");return -1;}if (des_file == NULL){fclose(src_file); return -1; }// read wav file header, copy it to the output file. fread(&wav_header, sizeof(wav_header), 1, src_file);fwrite(&wav_header, sizeof(wav_header), 1, des_file);// lets jump to the "data" chunk of the WAV file.if (strncmp(wav_header.datachunk_id, "data", 4)){wav_chunk_t chunk = { .size= wav_header.datachunk_size};// find the 'data' chunkdo {char* buf = malloc(chunk.size);fread(buf, chunk.size, 1, src_file);fwrite(buf, chunk.size, 1, des_file);free(buf);fread(&chunk, sizeof(wav_chunk_t), 1, src_file);fwrite(&chunk, sizeof(wav_chunk_t), 1, des_file);} while (strncmp(chunk.id, "data", 4));}// NNoM modelnnom_model_t *model = model = nnom_model_create();// 26 features, 0 offset, 26 bands, 512fft, 0 preempha, attached_energy_to_band0mfcc_t * mfcc = mfcc_create(NUM_FEATURES, 0, NUM_FEATURES, 512, 0, true);printf("\nProcessing file: %s\n", input_file);while(1) {// move buffer (50%) overlapping, move later 50% to the first 50, then fill memcpy(audio_buffer_16bit, &audio_buffer_16bit[AUDIO_FRAME_LEN/2], AUDIO_FRAME_LEN/2*sizeof(int16_t));// now read the new datasize = fread(&audio_buffer_16bit[AUDIO_FRAME_LEN / 2], AUDIO_FRAME_LEN / 2 * sizeof(int16_t), 1, src_file);if(size == 0)break;// get mfccmfcc_compute(mfcc, audio_buffer_16bit, mfcc_feature);//log_values(mfcc_feature, NUM_FEATURES, flog);// get the first and second derivative of mfccfor(uint32_t i=0; i< NUM_FEATURES; i++){mfcc_feature_diff[i] = mfcc_feature[i] - mfcc_feature_prev[i];mfcc_feature_diff1[i] = mfcc_feature_diff[i] - mfcc_feature_diff_prev[i];}memcpy(mfcc_feature_prev, mfcc_feature, NUM_FEATURES * sizeof(float));memcpy(mfcc_feature_diff_prev, mfcc_feature_diff, NUM_FEATURES * sizeof(float));// combine MFCC with derivatives memcpy(nn_features, mfcc_feature, NUM_FEATURES*sizeof(float));memcpy(&nn_features[NUM_FEATURES], mfcc_feature_diff, 10*sizeof(float));memcpy(&nn_features[NUM_FEATURES+10], mfcc_feature_diff1, 10*sizeof(float));//log_values(nn_features, NUM_FEATURES+20, flog);// quantise them using the same scale as training data (in keras), by 2^n. quantize_data(nn_features, nn_features_q7, NUM_FEATURES+20, 3);// run the mode with the new inputmemcpy(nnom_input_data, nn_features_q7, sizeof(nnom_input_data));model_run(model);// read the result, convert it back to float (q0.7 to float)for(int i=0; i< NUM_FEATURES; i++)band_gains[i] = (float)(nnom_output_data[i]) / 127.f;log_values(band_gains, NUM_FILTER, flog);// one more step, limit the change of gians, to smooth the speech, per RNNoise paperfor(int i=0; i< NUM_FEATURES; i++)band_gains[i] = _MAX(band_gains_prev[i]*0.8f, band_gains[i]); memcpy(band_gains_prev, band_gains, NUM_FEATURES *sizeof(float));// apply the dynamic gains to each frequency band. set_gains((float*)coeff_b, (float*)b_, band_gains, NUM_FILTER, NUM_ORDER);// convert 16bit to float for equalizerfor (int i = 0; i < AUDIO_FRAME_LEN/2; i++)audio_buffer[i] = audio_buffer_16bit[i + AUDIO_FRAME_LEN / 2] / 32768.f;// finally, we apply the equalizer to this audio frame to denoiseequalizer(audio_buffer, &audio_buffer[AUDIO_FRAME_LEN / 2], AUDIO_FRAME_LEN/2, (float*)b_,(float*)coeff_a, NUM_FILTER, NUM_ORDER);// convert the filtered signal back to int16for (int i = 0; i < AUDIO_FRAME_LEN / 2; i++)audio_buffer_filtered[i] = audio_buffer[i + AUDIO_FRAME_LEN / 2] * 32768.f *0.6f; // write the filtered frame to WAV file. fwrite(audio_buffer_filtered, 256*sizeof(int16_t), 1, des_file);}// print some model infomodel_io_format(model);model_stat(model);model_delete(model);fclose(flog);fclose(src_file);fclose(des_file);printf("\nNoisy signal '%s' has been de-noised by NNoM.\nThe output is saved to '%s'.\n", input_file, output_file);return 0;
}去掉wav的信息就能解析pcm了
创建cmake 文件 编译dll
cmake_minimum_required(VERSION 3.10)project(RnnDenoise)set(CMAKE_CXX_STANDARD 11)
set(CMAKE_CXX_STANDARD_REQUIRED True)
include_directories(nnom-master/port/)
include_directories(nnom-master/inc/)
include_directories(nnom-master/examples/rnn-denoise/)
include_directories(D:/java/jdk1.8x64/include/)
include_directories(D:/java/jdk1.8x64/include/win32/)
set(EXPLICIT_SOURCESRnnDenoise.cnnom-master/examples/rnn-denoise/mfcc.c
)
file(GLOB_RECURSE SRC_FILES "nnom-master/src/*/*.c")
set(SOURCES ${EXPLICIT_SOURCES} ${SRC_FILES})add_library(RnnDenoise SHARED ${SOURCES})
java 库封装示例
package com.lilin.demoasr.nnom;public class RnnDenoise implements AutoCloseable{private long rnndenoise;public long getRnndenoise() {return rnndenoise;}public void setRnndenoise(long rnndenoise) {this.rnndenoise = rnndenoise;}private static final Object globalLock = new Object();/***https://github.com/majianjia/nnom/blob/master/examples/rnn-denoise/**** @throws Exception*/public RnnDenoise() throws Exception {synchronized (globalLock) {RnnLoad.load("RnnDenoise");}this.rnndenoise = createRnnDenoise0();}private static native long createRnnDenoise0();private native short[] denoise0(long rnndenoise,short[] var1);/*** 固定320 每次 可以修改c 改大* @param input* @return*/public short[] denoise(short[] input) {// synchronized (this) {return this.denoise0(this.rnndenoise ,input);// }}private native long destroyRnnDenoise0();public void close() {synchronized (this) {this.destroyRnnDenoise0();this.rnndenoise = 0L;}}public boolean isClosed() {synchronized (this) {return this.rnndenoise == 0L;}}
}
test:
public static void main (String[] args) {String sList []= new String[]{"G:\\work\\ai\\ZipEnhancer\\r1.pcm","C:\\Users\\\\lilin\\Desktop\\16k.pcm"};// String sList []= new String[]{"C:\\Users\\\\lilin\\Desktop\\16k.pcm"};List< Thread> lts= new ArrayList<>();for (int i = 0; i < sList.length; i++) {String file =sList[i];int finalI = i;lts.add(new Thread(new Runnable() {@Overridepublic void run() {try {RnnDenoise rnnDenoise = new RnnDenoise();System.out.println(rnnDenoise.getRnndenoise());FileInputStream f = new FileInputStream(file);FileOutputStream f1 = new FileOutputStream("C:\\Users\\\\lilin\\Desktop\\"+ finalI +".pcm");int n=0;byte[] z = new byte[640];while ((n = f.read(z)) != -1) {short [] sa = new short[320];ByteBuffer.wrap(z).order(ByteOrder.LITTLE_ENDIAN).asShortBuffer().get(sa);short[] denoisedAudio = rnnDenoise.denoise(sa);byte[] z1 = new byte[640];ByteBuffer.wrap(z1).order(ByteOrder.LITTLE_ENDIAN).asShortBuffer().put(denoisedAudio);f1.write(z1);}System.out.println(finalI+"end.");rnnDenoise.close();f1.close();}catch (Exception e){e.printStackTrace();}}}));}for (Thread y: lts ) {y.start();}for (Thread y: lts ) {try{ y.join();}catch (Exception e){e.printStackTrace();}}System.out.println("end...");}
}nnom 默认的denoise_weights.h 是单例的无法同时创建多个实例 所以java无法在多线程使用， 可以自己更改下 主要涉及static变量和nnom_tensor_t 需要改用malloc的方式创建。
测试速度挺快的 ，几十分钟的很快降噪完成 ，也可以和freeswitch对接多路实时降噪 在识别，
如果模块或流程觉得麻烦可以到
https://item.taobao.com/item.htm?id=653611115230
视频教程yuexiajiayan的个人空间-yuexiajiayan个人主页-哔哩哔哩视频