前几天无意间看到一个项目rnnoise。
项目地址: https://github.com/xiph/rnnoise
基于RNN的音频降噪算法。
采用的是 GRU/LSTM 模型。
阅读下训练代码，可惜的是作者没有提供数据训练集。
不过基本可以断定他采用的数据集里，肯定有urbansound8k。
urbansound8k 数据集地址:
https://serv.cusp.nyu.edu/projects/urbansounddataset/urbansound8k.html
也可以考虑采用用作者训练的模型来构建数据集的做法，不过即费事，也麻烦。
经过实测，降噪效果很不错，特别是在背景噪声比较严重的情况下。
不过作者仅仅提供 pcm 的代码示例，并且还只支持48K采样率，
( 明显是为了兼容其另一个 项目 opus)
在很多应用场景下，这很不方便。
尽管稍微有点麻烦，但是事在人为，花了点时间，稍作修改。
具体修改如下：
1.支持wav格式
采用dr_wav(https://github.com/mackron/dr_libs/blob/master/dr_wav.h)
2.支持全部采样率
采样率的处理问题，采用简单粗暴法，
详情请移步博主另一篇小文《简洁明了的插值音频重采样算法例子 (附完整C代码)》
3.增加CMake文件
4.增加测试用 示例音频sample.wav
取自(https://github.com/orctom/rnnoise-java)
贴上完整示例代码 ：
#include <stdio.h>
#include "rnnoise.h"
#include <stdlib.h>
#include <stdint.h>
#define DR_WAV_IMPLEMENTATION
#include "dr_wav.h"
void wavWrite_int16(char *filename, int16_t *buffer, int sampleRate, uint32_t totalSampleCount) {
drwav_data_format format;
format.container = drwav_container_riff;
format.format = DR_WAVE_FORMAT_PCM;
format.channels = ;
format.sampleRate = (drwav_uint32) sampleRate;
format.bitsPerSample = ;
drwav *pWav = drwav_open_file_write(filename, &format);
if (pWav) {
drwav_uint64 samplesWritten = drwav_write(pWav, totalSampleCount, buffer);
drwav_uninit(pWav);
if (samplesWritten != totalSampleCount) {
fprintf(stderr, "ERROR\n");
exit();
}
}
}
int16_t *wavRead_int16(char *filename, uint32_t *sampleRate, uint64_t *totalSampleCount) {
unsigned int channels;
int16_t *buffer = drwav_open_and_read_file_s16(filename, &channels, sampleRate, totalSampleCount);
if (buffer == NULL) {
fprintf(stderr, "ERROR\n");
exit();
}
) {
drwav_free(buffer);
buffer = NULL;
*sampleRate = ;
*totalSampleCount = ;
}
return buffer;
}
void splitpath(const char *path, char *drv, char *dir, char *name, char *ext) {
const char *end;
const char *p;
const char *s;
] && path[] == ':') {
if (drv) {
*drv++ = *path++;
*drv++ = *path++;
*drv = '\0';
}
} else if (drv)
*drv = '\0';
for (end = path; *end && *end != ':';)
end++;
for (p = end; p > path && *--p != '\\' && *p != '/';)
if (*p == '.') {
end = p;
break;
}
if (ext)
for (s = end; (*ext = *s++);)
ext++;
for (p = end; p > path;)
if (*--p == '\\' || *p == '/') {
p++;
break;
}
if (name) {
for (s = p; s < end;)
*name++ = *s++;
*name = '\0';
}
if (dir) {
for (s = path; s < p;)
*dir++ = *s++;
*dir = '\0';
}
}
void resampleData(const int16_t *sourceData, int32_t sampleRate, uint32_t srcSize, int16_t *destinationData,
int32_t newSampleRate) {
if (sampleRate == newSampleRate) {
memcpy(destinationData, sourceData, srcSize * sizeof(int16_t));
return;
}
uint32_t last_pos = srcSize - ;
uint32_t dstSize = (uint32_t) (srcSize * ((float) newSampleRate / sampleRate));
; idx < dstSize; idx++) {
float index = ((float) idx * sampleRate) / (newSampleRate);
uint32_t p1 = (uint32_t) index;
float coef = index - p1;
uint32_t p2 = (p1 == last_pos) ? last_pos : p1 + ;
destinationData[idx] = (int16_t) ((1.0f - coef) * sourceData[p1] + coef * sourceData[p2]);
}
}
void f32_to_s16(int16_t *pOut, const float *pIn, size_t sampleCount) {
if (pOut == NULL || pIn == NULL) {
return;
}
; i < sampleCount; ++i) {
*pOut++ = (short) pIn[i];
}
}
void s16_to_f32(float *pOut, const int16_t *pIn, size_t sampleCount) {
if (pOut == NULL || pIn == NULL) {
return;
}
; i < sampleCount; ++i) {
*pOut++ = pIn[i];
}
}
void denoise_proc(int16_t *buffer, uint32_t buffen_len) {
;
DenoiseState *st;
st = rnnoise_create();
float patch_buffer[frame_size];
if (st != NULL) {
uint32_t frames = buffen_len / frame_size;
uint32_t lastFrame = buffen_len % frame_size;
; i < frames; ++i) {
s16_to_f32(patch_buffer, buffer, frame_size);
rnnoise_process_frame(st, patch_buffer, patch_buffer);
f32_to_s16(buffer, patch_buffer, frame_size);
buffer += frame_size;
}
) {
memset(patch_buffer, , frame_size * sizeof(float));
s16_to_f32(patch_buffer, buffer, lastFrame);
rnnoise_process_frame(st, patch_buffer, patch_buffer);
f32_to_s16(buffer, patch_buffer, lastFrame);
}
}
rnnoise_destroy(st);
}
void rnnDeNoise(char *in_file, char *out_file) {
uint32_t in_sampleRate = ;
uint64_t in_size = ;
int16_t *data_in = wavRead_int16(in_file, &in_sampleRate, &in_size);
uint32_t out_sampleRate = ;
uint32_t out_size = (uint32_t) (in_size * ((float) out_sampleRate / in_sampleRate));
int16_t *data_out = (int16_t *) malloc(out_size * sizeof(int16_t));
if (data_in != NULL && data_out != NULL) {
resampleData(data_in, in_sampleRate, (uint32_t) in_size, data_out, out_sampleRate);
denoise_proc(data_out, out_size);
resampleData(data_out, out_sampleRate, (uint32_t) out_size, data_in, in_sampleRate);
wavWrite_int16(out_file, data_in, in_sampleRate, (uint32_t) in_size);
free(data_in);
free(data_out);
} else {
if (data_in) free(data_in);
if (data_out) free(data_out);
}
}
int main(int argc, char **argv) {
printf("Audio Noise Reduction\n");
printf("blog:http://tntmonks.cnblogs.com/\n");
printf("e-mail:gaozhihan@vip.qq.com\n");
)
;
];
];
];
];
];
];
splitpath(in_file, drive, dir, fname, ext);
sprintf(out_file, "%s%s%s_out%s", drive, dir, fname, ext);
rnnDeNoise(in_file, out_file);
printf("press any key to exit.\n");
getchar();
;
}
不多写注释，直接看代码吧。
项目地址：https://github.com/cpuimage/rnnoise
示例具体流程为：
加载wav(拖放wav文件到可执行文件上)->重采样降噪->保存wav
若有其他相关问题或者需求也可以邮件联系俺探讨。
邮箱地址是:
gaozhihan@vip.qq.com