浏览器中的mp3流解码

mp3 stream decoding in browser

本文关键字:解码 mp3 浏览器      更新时间:2023-10-16

我正在尝试使用emscripten和libmad在浏览器中设置一个mp3流接收器
我设法用低级api解码mp3文件,将其完全加载到内存中。我的下一步是将其分块加载
在给定的例子中,我模拟了具有随机大小(从20到40k字节)的分配缓冲区的碎片包,并将文件逐部分复制到这些缓冲区中。

我的解码算法与这个问题的答案相关,但有点不同。主要对象是Decoder,它通过addFragment方法接收片段。解码器有一个挂起的片段拉和一个粘合缓冲区。当用户添加第一个片段时,它的尾部被复制到粘合缓冲区的前半部分。当添加第二个片段时,它开始被复制到胶水的后半部分。当解码器到达活动缓冲区末端时,它切换到粘合,反之亦然。我确保所有这些缓冲区部分是一致的,并且mad_stream指向切换前所指向的同一逻辑字节。

解码器.cpp的重要片段

void Decoder::addFragment          //adds the fragment to decoding queue
(intptr_t bufferPtr, uint32_t length)
{
if (length < GLUE_LENGTH / 2) {
return;
}
uint8_t* buffer = (uint8_t(*))bufferPtr;
RawBuffer rb = {buffer, length};
pending.push_back(rb);
switch (state) {
case empty:
mad_stream_buffer(&stream, buffer, length);
for (int i = 0; i < GLUE_LENGTH/2; ++i) {
glue[i] = buffer[length - GLUE_LENGTH/2 + i];
}
state = onBufferHalf;
prepareNextBuffer();
break;
case onBufferHalf:
for (int i = 0; i < GLUE_LENGTH/2; ++i) {
glue[GLUE_LENGTH/2 + i] = buffer[i];
}
state = onBufferFull;
break;
case onGlueHalf:
for (int i = 0; i < GLUE_LENGTH/2; ++i) {
glue[GLUE_LENGTH/2 + i] = buffer[i];
}
state = onGlueFull;
cached = false;
prepareNextBuffer();
break;
default:
break;
}
}
emscripten::val Decoder::decode     //decodes up to requested amount of frames
(uint32_t count)
{
emscripten::val ret = emscripten::val::undefined();
int available = framesLeft(count);
if (available > 0) {
ret = context.call<emscripten::val>("createBuffer", channels, available * samplesPerFrame, sampleRate);
std::vector<emscripten::val> chans(channels, emscripten::val::undefined());
for (int i = 0; i < channels; ++i) {
chans[i] = ret.call<emscripten::val>("getChannelData", i);
} 
for (int i = 0; i < available; ++i) {
int res = mad_frame_decode(&frame, &stream);
if (res != 0) {
if (MAD_RECOVERABLE(stream.error)) {
continue;
} else {
break;
}
}
mad_synth_frame(&synth, &frame);
for (int j = 0; j < samplesPerFrame; ++j) {
for (int k = 0; k < channels; ++k) {
float value = mad_f_todouble(synth.pcm.samples[k][j]);
chans[k].set(std::to_string(success * samplesPerFrame + j), emscripten::val(value));
}
}
}
cachedLength -= available;
if (cachedLength == 0) {
cached = false;
prepareNextBuffer();
}
}
return ret;
}

//tells how many frames can be decoded on the same
//sample rate, same amount of channels without switching the buffers
//it is required in Decoder::decode method to understand the size of 
//allocating AudioContext::AudioBuffer.
uint32_t Decoder::framesLeft(uint32_t max)
{
if (state == empty || state == onGlueHalf) {
return 0;
}
if (cached == false) {
mad_stream probe;
mad_header ph;
initializeProbe(probe);
mad_header_init(&ph);
while (cachedLength < max) {
if (mad_header_decode(&ph, &probe) == 0) {
if (sampleRate == 0) {
sampleRate = ph.samplerate;
channels = MAD_NCHANNELS(&ph);
samplesPerFrame = MAD_NSBSAMPLES(&ph) * 32;
} else {
if (sampleRate != ph.samplerate || channels != MAD_NCHANNELS(&ph) || samplesPerFrame != MAD_NSBSAMPLES(&ph) * 32) {
break;
}
}
if (probe.next_frame > probe.this_frame) {
++cachedLength;
}
} else {
if (!MAD_RECOVERABLE(probe.error)) {
break;
}
}
}
cachedNext = probe.next_frame;
cachedThis = probe.this_frame;
cachedError = probe.error;
mad_header_finish(&ph);
mad_stream_finish(&probe);
cached = true;
}
return std::min(cachedLength, max);
}
//this method fastforwards the stream
//to the cached end
void Decoder::pullBuffer()
{
if (cached == false) {
throw 2;
}
stream.this_frame = cachedThis;
stream.next_frame = cachedNext;
stream.error = cachedError;
}
//this method switches the stream to glue buffer
//or to the next pending buffer
//copies the parts to the glue buffer if required
void Decoder::changeBuffer()
{
uint32_t left;
switch (state) {
case empty:
throw 3;
case onBufferHalf:
switchToGlue();
state = onGlueHalf;
break;
case onBufferFull:
switchToGlue();
state = onGlueFull;
break;
case onGlueHalf:
throw 4;
break;
case onGlueFull:
switchBuffer(pending[0].ptr, pending[0].length);
for (int i = 0; i < GLUE_LENGTH/2; ++i) {
glue[i] = pending[0].ptr[pending[0].length - GLUE_LENGTH/2 + i];
}
state = onBufferHalf;
if (pending.size() > 1) {
for (int i = 0; i < GLUE_LENGTH/2; ++i) {
glue[GLUE_LENGTH/2 + i] = pending[1].ptr[i];
}
state = onBufferFull;
}
}
cached = false;
}
//this method seeks the decodable data in pending buffers
//prepares if any proper data has been found
void Decoder::prepareNextBuffer()
{
bool shift;
do {
shift = false;
framesLeft();
if (cachedLength == 0 && state != empty && state != onGlueHalf) {
pullBuffer();
changeBuffer();
shift = true;
}
} while (shift);
}
//low level method to switch to glue buffer, also frees the drained fragment
void Decoder::switchToGlue()
{
switchBuffer(glue, GLUE_LENGTH);
stream.error = MAD_ERROR_NONE;
free(pending[0].ptr);
pending.pop_front();
}
//low level method which actually switch mad_stream
//to another buffer
void Decoder::switchBuffer(uint8_t* bufferPtr, uint32_t length)
{
uint32_t left;
left = stream.bufend - stream.next_frame;
mad_stream_buffer(&stream, bufferPtr + GLUE_LENGTH / 2 - left, length - (GLUE_LENGTH / 2 - left));
stream.error = MAD_ERROR_NONE;
}

这是我的回购完整代码。要尝试它,你需要用CMake构建它(应该安装emscripten),并从浏览器的构建目录中打开index.html。

问题
播放失真。我试着在shift之前和之后检查最后一个成功帧周围的字节,mad_stream的所有不同子结构——一切似乎都正常工作,但仍然没有。我的最新进展是在这里构建和托管的。我真的被卡住了,我不知道该怎么做才能消除播放中的失真。

如果有人帮我,我将不胜感激。

我找到了!MAD工作得很好,只是因为我的内部计数器,我一直在跳过输出中的第一个解码帧。

for (int i = 0; success < available; ++i) {
int res = mad_frame_decode(frame, stream);
if (res == 0) {
++**success**;
} else {
if (MAD_RECOVERABLE(stream->error)) {
std::cout << "Unexpected error during the decoding process: " << mad_stream_errorstr(stream) << std::endl;
continue;
} else {
break;
}
}
mad_synth_frame(synth, frame);
for (int j = 0; j < samplesPerFrame; ++j) {
for (int k = 0; k < channels; ++k) {
float value = mad_f_todouble(synth->pcm.samples[k][j]);
chans[k].set(std::to_string(success * samplesPerFrame + j), emscripten::val(value));
}
}
}

success更改为i,效果良好。

这里有多个不同的答案,但有一个解决方案是通过socket.io。这将避免在服务器上进行分割,并委托客户端为您进行分割。

这里缺少的部分是通过mp3解码器运行二进制文件。有些库可以自动确定格式,但您可能还必须传递编码类型,以便它知道如何解析二进制流。它可能已经在极光中提供了,但我不熟悉。