总结网页音频直播的方案和遇到的问题。
代码:(github,待整理)
结果: 使用opus音频编码,web audio api 播放,可以达到100ms以内延时,高质量,低流量的音频直播。
背景: VDI(虚拟桌面) h264网页版预研,继h264视频直播方案解决之后的又一个对延时有高要求的音频直播方案(交互性,音视频同步)。
前提: flexVDI开源项目对音频的支持只实现了对未编码压缩的PCM音频数据。并且效果不好,要么卡顿,要么延时,流量在2~3Mbps(根据缓冲的大小)。
解决方案: 在spice server端对音频采用opus进行编码,flexVDI playback通道拿到opus packet数据后,调用opus js解码库解码成PCM数据,喂给audioContext进行播放。
流程简介:flexVDI palyback通道接收opus音频数据,调用libopus.js解码得到PCM数据,保存到buffer。创建scriptProcessorNode, 在onaudioprocess函数中从buffer里面拿到PCM数据,
按声道填充outputBuffer, 把scriptProcessorNode连接到audioContext.destination进行播放。具体代码见后文或者github。
opus编解码接口介绍:
参考:http://opus-codec.org/docs/opus_api-1.2/index.html
一、下面是我用opus c库解码opus音频,再用ffplay播放PCM数据的一个demo,可以看看opus解码接口是怎么使用的:
#include <stdio.h> #include <stdlib.h> #include <string.h> #include "opus.h" /* static void int_to_char(opus_uint32 i, unsigned char ch[4]) { ch[0] = i>>24; ch[1] = (i>>16)&0xFF; ch[2] = (i>>8)&0xFF; ch[3] = i&0xFF; }*/ static opus_uint32 char_to_int(unsigned char ch[4]) { return ((opus_uint32)ch[0]<<24) | ((opus_uint32)ch[1]<<16) | ((opus_uint32)ch[2]<< 8) | (opus_uint32)ch[3]; } int main(int argc, char** argv) { opus_int32 sampleRate = 0; int channels = 0, err = 0, len = 0; int max_payload_bytes = 1500; int max_frame_size = 48000*2; OpusDecoder* dec = NULL; sampleRate = (opus_int32)atol(argv[1]); channels = atoi(argv[2]); FILE* fin = fopen(argv[3], "rb"); FILE* fout = fopen(argv[4], "wb+"); short *out; unsigned char* fbytes, *data; //in = (short*)malloc(max_frame_size*channels*sizeof(short)); out = (short*)malloc(max_frame_size*channels*sizeof(short)); /* We need to allocate for 16-bit PCM data, but we store it as unsigned char. */ fbytes = (unsigned char*)malloc(max_frame_size*channels*sizeof(short)); data = (unsigned char*)calloc(max_payload_bytes, sizeof(unsigned char)); dec = opus_decoder_create(sampleRate, channels, &err); int nBytesRead = 0; opus_uint64 tot_out = 0; while(1){ unsigned char ch[4] = {0}; nBytesRead = fread(ch, 1, 4, fin); if(nBytesRead != 4) break; len = char_to_int(ch); nBytesRead = fread(data, 1, len, fin); if(nBytesRead != len) break; opus_int32 output_samples = max_frame_size; output_samples = opus_decode(dec, data, len, out, output_samples, 0); int i; for(i=0; i < output_samples*channels; i++) { short s; s=out[i]; fbytes[2*i]=s&0xFF; fbytes[2*i+1]=(s>>8)&0xFF; } if (fwrite(fbytes, sizeof(short)*channels, output_samples, fout) != (unsigned)output_samples){ fprintf(stderr, "Error writing.\n"); return EXIT_FAILURE; } tot_out += output_samples; } printf("tot_out: %llu \n", tot_out); return 0; }
这个程序对opus packets组成的文件(简单的length+packet格式)解码后得到PCM数据,再用ffplay播放PCM数据,看能否正常播放:
ffplay -f f32le -ac 1 -ar 48000 input_audio // 播放float32型PCM数据
ffplay -f s16le -ac 1 -ar 48000 input_audio //播放short16型PCM数据
ac表示声道数, ar表示采样率, input_audio是PCM音频文件。
二、要获取PCM数据文件,首先要得到opus packet二进制文件, 所以这里涉及到浏览器如何保存二进制文件到本地的问题:
参考代码:
var saveFile = (function(){ var a = document.createElement("a"); document.body.appendChild(a); a.style = "display:none"; return function(data, name){ var blob = new Blob([data]); var url = window.URL.createObjectURL(blob); a.href = url; a.download = name; a.click(); window.URL.revokeObjectURL(url); }; }()); saveFile(data, ‘test.pcm‘);
说明:首先把二进制数据写到typedArray中,然后用这个buffer构造Blob对象,生成URL, 再使用a标签把这个blob下载到本地。
三、利用audioContext播放PCM音频数据的两种方案:
(1)flexVDI的实现
参考:https://github.com/flexVDI/spice-web-client
function play(buffer, dataTimestamp) { // Each data packet is 16 bits, the first being left channel data and the second being right channel data (LR-LR-LR-LR...) //var audio = new Int16Array(buffer); var audio = new Float32Array(buffer); // We split the audio buffer in two channels. Float32Array is the type required by Web Audio API var left = new Float32Array(audio.length / 2); var right = new Float32Array(audio.length / 2); var channelCounter = 0; var audioContext = this.audioContext; var len = audio.length; for (var i = 0; i < len; ) { //because the audio data spice gives us is 16 bits signed int (32768) and we wont to get a float out of it (between -1.0 and 1.0) left[channelCounter] = audio[i++] / 32768; right[channelCounter] = audio[i++] / 32768; channelCounter++; } var source = audioContext[‘createBufferSource‘](); // creates a sound source var audioBuffer = audioContext[‘createBuffer‘](2, channelCounter, this.frequency); audioBuffer[‘getChannelData‘](0)[‘set‘](left); audioBuffer[‘getChannelData‘](1)[‘set‘](right); source[‘buffer‘] = audioBuffer; source[‘connect‘](this.audioContext[‘destination‘]); source[‘start‘](0); }
注: buffer中保存的是short 型PCM数据,这里为了简单,去掉了对时间戳的处理,因为source.start(0)表示立即播放。如果是float型数据,不需要除以32768.
(2)ws-audio-api的实现
参考:https://github.com/Ivan-Feofanov/ws-audio-api
var bufL = new Float32Array(this.config.codec.bufferSize); var bufR = new Float32Array(this.config.codec.bufferSize); this.scriptNode = audioContext.createScriptProcessor(this.config.codec.bufferSize, 0, 2); if (typeof AudioBuffer.prototype.copyToChannel === "function") { this.scriptNode.onaudioprocess = function(e) { var buf = e.outputBuffer; _this.process(bufL, bufR); //获取PCM数据到bufL, bufR buf.copyToChannel(bufL, 0); buf.copyToChannel(bufR, 1); }; } else { this.scriptNode.onaudioprocess = function(e) { var buf = e.outputBuffer; _this.process(bufL, bufR); buf.getChannelData(0).set(bufL); buf.getChannelData(1).set(bufR); }; } this.scriptNode.connect(audioContext.destination);
延时卡顿的问题:audioContext有的浏览器默认是48000采样率,有的浏览器默认是44100的采样率,如果喂给audioContext的PCM数据的采样率不匹配,就会产生延时和卡顿的问题。