From ddfcd68c803c226f4b2a60b0666b29c24d04c259 Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 14 四月 2023 13:15:29 +0800
Subject: [PATCH] Merge pull request #353 from alibaba-damo-academy/dev_lyb
---
funasr/runtime/onnxruntime/src/Audio.cpp | 262 ++++++++++++++++++++++++++++++++++++---------------
1 files changed, 183 insertions(+), 79 deletions(-)
diff --git a/funasr/runtime/onnxruntime/src/Audio.cpp b/funasr/runtime/onnxruntime/src/Audio.cpp
index bce3a90..38b6de8 100644
--- a/funasr/runtime/onnxruntime/src/Audio.cpp
+++ b/funasr/runtime/onnxruntime/src/Audio.cpp
@@ -3,10 +3,95 @@
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
+#include <fstream>
+#include <assert.h>
#include "Audio.h"
+#include "precomp.h"
using namespace std;
+
+// see http://soundfile.sapp.org/doc/WaveFormat/
+// Note: We assume little endian here
+struct WaveHeader {
+ bool Validate() const {
+ // F F I R
+ if (chunk_id != 0x46464952) {
+ printf("Expected chunk_id RIFF. Given: 0x%08x\n", chunk_id);
+ return false;
+ }
+ // E V A W
+ if (format != 0x45564157) {
+ printf("Expected format WAVE. Given: 0x%08x\n", format);
+ return false;
+ }
+
+ if (subchunk1_id != 0x20746d66) {
+ printf("Expected subchunk1_id 0x20746d66. Given: 0x%08x\n",
+ subchunk1_id);
+ return false;
+ }
+
+ if (subchunk1_size != 16) { // 16 for PCM
+ printf("Expected subchunk1_size 16. Given: %d\n",
+ subchunk1_size);
+ return false;
+ }
+
+ if (audio_format != 1) { // 1 for PCM
+ printf("Expected audio_format 1. Given: %d\n", audio_format);
+ return false;
+ }
+
+ if (num_channels != 1) { // we support only single channel for now
+ printf("Expected single channel. Given: %d\n", num_channels);
+ return false;
+ }
+ if (byte_rate != (sample_rate * num_channels * bits_per_sample / 8)) {
+ return false;
+ }
+
+ if (block_align != (num_channels * bits_per_sample / 8)) {
+ return false;
+ }
+
+ if (bits_per_sample != 16) { // we support only 16 bits per sample
+ printf("Expected bits_per_sample 16. Given: %d\n",
+ bits_per_sample);
+ return false;
+ }
+ return true;
+ }
+
+ // See https://en.wikipedia.org/wiki/WAV#Metadata and
+ // https://www.robotplanet.dk/audio/wav_meta_data/riff_mci.pdf
+ void SeekToDataChunk(std::istream &is) {
+ // a t a d
+ while (is && subchunk2_id != 0x61746164) {
+ // const char *p = reinterpret_cast<const char *>(&subchunk2_id);
+ // printf("Skip chunk (%x): %c%c%c%c of size: %d\n", subchunk2_id, p[0],
+ // p[1], p[2], p[3], subchunk2_size);
+ is.seekg(subchunk2_size, std::istream::cur);
+ is.read(reinterpret_cast<char *>(&subchunk2_id), sizeof(int32_t));
+ is.read(reinterpret_cast<char *>(&subchunk2_size), sizeof(int32_t));
+ }
+ }
+
+ int32_t chunk_id;
+ int32_t chunk_size;
+ int32_t format;
+ int32_t subchunk1_id;
+ int32_t subchunk1_size;
+ int16_t audio_format;
+ int16_t num_channels;
+ int32_t sample_rate;
+ int32_t byte_rate;
+ int16_t block_align;
+ int16_t bits_per_sample;
+ int32_t subchunk2_id; // a tag of this chunk
+ int32_t subchunk2_size; // size of subchunk2
+};
+static_assert(sizeof(WaveHeader) == WAV_HEADER_SIZE, "");
class AudioWindow {
private:
@@ -56,7 +141,7 @@
float frame_length = 400;
float frame_shift = 160;
float num_new_samples =
- ceil((num_samples - 400) / frame_shift) * frame_shift + frame_length;
+ ceil((num_samples - frame_length) / frame_shift) * frame_shift + frame_length;
end = start + num_new_samples;
len = (int)num_new_samples;
@@ -111,62 +196,95 @@
void Audio::disp()
{
- printf("Audio time is %f s. len is %d\n", (float)speech_len / 16000,
+ printf("Audio time is %f s. len is %d\n", (float)speech_len / model_sample_rate,
speech_len);
}
float Audio::get_time_len()
{
- return (float)speech_len / 16000;
- //speech_len);
+ return (float)speech_len / model_sample_rate;
}
-bool Audio::loadwav(const char *filename)
+void Audio::wavResample(int32_t sampling_rate, const float *waveform,
+ int32_t n)
{
+ printf(
+ "Creating a resampler:\n"
+ " in_sample_rate: %d\n"
+ " output_sample_rate: %d\n",
+ sampling_rate, static_cast<int32_t>(model_sample_rate));
+ float min_freq =
+ std::min<int32_t>(sampling_rate, model_sample_rate);
+ float lowpass_cutoff = 0.99 * 0.5 * min_freq;
+ int32_t lowpass_filter_width = 6;
+ //FIXME
+ //auto resampler = new LinearResample(
+ // sampling_rate, model_sample_rate, lowpass_cutoff, lowpass_filter_width);
+ auto resampler = std::make_unique<LinearResample>(
+ sampling_rate, model_sample_rate, lowpass_cutoff, lowpass_filter_width);
+ std::vector<float> samples;
+ resampler->Resample(waveform, n, true, &samples);
+ //reset speech_data
+ speech_len = samples.size();
+ if (speech_data != NULL) {
+ free(speech_data);
+ }
+ speech_data = (float*)malloc(sizeof(float) * speech_len);
+ memset(speech_data, 0, sizeof(float) * speech_len);
+ copy(samples.begin(), samples.end(), speech_data);
+}
+
+bool Audio::loadwav(const char *filename, int32_t* sampling_rate)
+{
+ WaveHeader header;
if (speech_data != NULL) {
free(speech_data);
}
if (speech_buff != NULL) {
free(speech_buff);
}
-
+
offset = 0;
-
- FILE *fp;
- fp = fopen(filename, "rb");
- if (fp == nullptr)
+ std::ifstream is(filename, std::ifstream::binary);
+ is.read(reinterpret_cast<char *>(&header), sizeof(header));
+ if(!is){
+ fprintf(stderr, "Failed to read %s\n", filename);
return false;
- fseek(fp, 0, SEEK_END); /*瀹氫綅鍒版枃浠舵湯灏�*/
- uint32_t nFileLen = ftell(fp); /*寰楀埌鏂囦欢澶у皬*/
- fseek(fp, 44, SEEK_SET); /*璺宠繃wav鏂囦欢澶�*/
-
- speech_len = (nFileLen - 44) / 2;
- speech_align_len = (int)(ceil((float)speech_len / align_size) * align_size);
- speech_buff = (int16_t *)malloc(sizeof(int16_t) * speech_align_len);
+ }
+
+ *sampling_rate = header.sample_rate;
+ // header.subchunk2_size contains the number of bytes in the data.
+ // As we assume each sample contains two bytes, so it is divided by 2 here
+ speech_len = header.subchunk2_size / 2;
+ speech_buff = (int16_t *)malloc(sizeof(int16_t) * speech_len);
if (speech_buff)
{
- memset(speech_buff, 0, sizeof(int16_t) * speech_align_len);
- int ret = fread(speech_buff, sizeof(int16_t), speech_len, fp);
- fclose(fp);
+ memset(speech_buff, 0, sizeof(int16_t) * speech_len);
+ is.read(reinterpret_cast<char *>(speech_buff), header.subchunk2_size);
+ if (!is) {
+ fprintf(stderr, "Failed to read %s\n", filename);
+ return false;
+ }
+ speech_data = (float*)malloc(sizeof(float) * speech_len);
+ memset(speech_data, 0, sizeof(float) * speech_len);
- speech_data = (float*)malloc(sizeof(float) * speech_align_len);
- memset(speech_data, 0, sizeof(float) * speech_align_len);
- int i;
float scale = 1;
-
if (data_type == 1) {
scale = 32768;
}
-
- for (i = 0; i < speech_len; i++) {
+ for (int32_t i = 0; i != speech_len; ++i) {
speech_data[i] = (float)speech_buff[i] / scale;
+ }
+
+ //resample
+ if(*sampling_rate != model_sample_rate){
+ wavResample(*sampling_rate, speech_data, speech_len);
}
AudioFrame* frame = new AudioFrame(speech_len);
frame_queue.push(frame);
-
return true;
}
@@ -174,57 +292,54 @@
return false;
}
-
-bool Audio::loadwav(const char* buf, int nFileLen)
+bool Audio::loadwav(const char* buf, int nFileLen, int32_t* sampling_rate)
{
-
-
-
+ WaveHeader header;
if (speech_data != NULL) {
free(speech_data);
}
if (speech_buff != NULL) {
free(speech_buff);
}
-
offset = 0;
- size_t nOffset = 0;
+ std::memcpy(&header, buf, sizeof(header));
-#define WAV_HEADER_SIZE 44
-
- speech_len = (nFileLen - WAV_HEADER_SIZE) / 2;
- speech_align_len = (int)(ceil((float)speech_len / align_size) * align_size);
- speech_buff = (int16_t*)malloc(sizeof(int16_t) * speech_align_len);
+ *sampling_rate = header.sample_rate;
+ speech_len = header.subchunk2_size / 2;
+ speech_buff = (int16_t *)malloc(sizeof(int16_t) * speech_len);
if (speech_buff)
{
- memset(speech_buff, 0, sizeof(int16_t) * speech_align_len);
+ memset(speech_buff, 0, sizeof(int16_t) * speech_len);
memcpy((void*)speech_buff, (const void*)(buf + WAV_HEADER_SIZE), speech_len * sizeof(int16_t));
+ speech_data = (float*)malloc(sizeof(float) * speech_len);
+ memset(speech_data, 0, sizeof(float) * speech_len);
- speech_data = (float*)malloc(sizeof(float) * speech_align_len);
- memset(speech_data, 0, sizeof(float) * speech_align_len);
- int i;
float scale = 1;
-
if (data_type == 1) {
scale = 32768;
}
- for (i = 0; i < speech_len; i++) {
+ for (int32_t i = 0; i != speech_len; ++i) {
speech_data[i] = (float)speech_buff[i] / scale;
}
+
+ //resample
+ if(*sampling_rate != model_sample_rate){
+ wavResample(*sampling_rate, speech_data, speech_len);
+ }
+ AudioFrame* frame = new AudioFrame(speech_len);
+ frame_queue.push(frame);
return true;
}
else
return false;
-
}
-
-bool Audio::loadpcmwav(const char* buf, int nBufLen)
+bool Audio::loadpcmwav(const char* buf, int nBufLen, int32_t* sampling_rate)
{
if (speech_data != NULL) {
free(speech_data);
@@ -234,32 +349,28 @@
}
offset = 0;
- size_t nOffset = 0;
-
-
-
speech_len = nBufLen / 2;
- speech_align_len = (int)(ceil((float)speech_len / align_size) * align_size);
- speech_buff = (int16_t*)malloc(sizeof(int16_t) * speech_align_len);
+ speech_buff = (int16_t*)malloc(sizeof(int16_t) * speech_len);
if (speech_buff)
{
- memset(speech_buff, 0, sizeof(int16_t) * speech_align_len);
+ memset(speech_buff, 0, sizeof(int16_t) * speech_len);
memcpy((void*)speech_buff, (const void*)buf, speech_len * sizeof(int16_t));
+ speech_data = (float*)malloc(sizeof(float) * speech_len);
+ memset(speech_data, 0, sizeof(float) * speech_len);
- speech_data = (float*)malloc(sizeof(float) * speech_align_len);
- memset(speech_data, 0, sizeof(float) * speech_align_len);
-
-
- int i;
float scale = 1;
-
if (data_type == 1) {
scale = 32768;
}
- for (i = 0; i < speech_len; i++) {
+ for (int32_t i = 0; i != speech_len; ++i) {
speech_data[i] = (float)speech_buff[i] / scale;
+ }
+
+ //resample
+ if(*sampling_rate != model_sample_rate){
+ wavResample(*sampling_rate, speech_data, speech_len);
}
AudioFrame* frame = new AudioFrame(speech_len);
@@ -269,13 +380,10 @@
}
else
return false;
-
-
}
-bool Audio::loadpcmwav(const char* filename)
+bool Audio::loadpcmwav(const char* filename, int32_t* sampling_rate)
{
-
if (speech_data != NULL) {
free(speech_data);
}
@@ -293,34 +401,31 @@
fseek(fp, 0, SEEK_SET);
speech_len = (nFileLen) / 2;
- speech_align_len = (int)(ceil((float)speech_len / align_size) * align_size);
- speech_buff = (int16_t*)malloc(sizeof(int16_t) * speech_align_len);
+ speech_buff = (int16_t*)malloc(sizeof(int16_t) * speech_len);
if (speech_buff)
{
- memset(speech_buff, 0, sizeof(int16_t) * speech_align_len);
+ memset(speech_buff, 0, sizeof(int16_t) * speech_len);
int ret = fread(speech_buff, sizeof(int16_t), speech_len, fp);
fclose(fp);
- speech_data = (float*)malloc(sizeof(float) * speech_align_len);
- memset(speech_data, 0, sizeof(float) * speech_align_len);
+ speech_data = (float*)malloc(sizeof(float) * speech_len);
+ memset(speech_data, 0, sizeof(float) * speech_len);
-
-
- int i;
float scale = 1;
-
if (data_type == 1) {
scale = 32768;
}
-
- for (i = 0; i < speech_len; i++) {
+ for (int32_t i = 0; i != speech_len; ++i) {
speech_data[i] = (float)speech_buff[i] / scale;
}
+ //resample
+ if(*sampling_rate != model_sample_rate){
+ wavResample(*sampling_rate, speech_data, speech_len);
+ }
AudioFrame* frame = new AudioFrame(speech_len);
frame_queue.push(frame);
-
return true;
}
@@ -328,7 +433,6 @@
return false;
}
-
int Audio::fetch_chunck(float *&dout, int len)
{
--
Gitblit v1.9.1