Files
mediabunny/packages/server/src/audio-encoder.ts
T

441 lines
15 KiB
TypeScript

/*!
* Copyright (c) 2026-present, Vanilagy and contributors
*
* This Source Code Form is subject to the terms of the Mozilla Public
* License, v. 2.0. If a copy of the MPL was not distributed with this
* file, You can obtain one at https://mozilla.org/MPL/2.0/.
*/
import {
AudioCodec,
AudioSample,
CustomAudioEncoder,
type MaybePromise,
Quality,
EncodedPacket,
} from 'mediabunny';
import * as NodeAv from 'node-av';
import { CODEC_TO_CODEC_ID, getChannelLayout, getDtsChannelLayout } from './misc';
import { assert, toUint8Array } from '../../../src/misc';
import { copyAudioSampleToAvFrame, AvFrameAudioSampleResource } from './audio-sample';
import {
AdtsHeaderTemplate,
buildAdtsHeaderTemplate,
parseAacAudioSpecificConfig,
} from '../../../shared/aac-misc';
const AAC_SAMPLE_RATES
= [96000, 88200, 64000, 48000, 44100, 32000, 24000, 22050, 16000, 12000, 11025, 8000, 7350];
const OPUS_SAMPLE_RATES = [8000, 12000, 16000, 24000, 48000];
const MP3_SAMPLE_RATES = [8000, 11025, 12000, 16000, 22050, 24000, 32000, 44100, 48000];
const AC3_SAMPLE_RATES = [32000, 44100, 48000];
const DTS_SAMPLE_RATES = [8000, 11025, 12000, 16000, 22050, 24000, 32000, 44100, 48000];
/** DTS only has layouts for these channel counts; 3 and 7 have no mapping. */
const DTS_CHANNEL_COUNTS = [1, 2, 4, 5, 6];
const DTS_MAX_FRAME_SIZE = 16384;
const FRAME_SIZE_FALLBACK = 1024; // Just 'cause
export class NodeAvAudioEncoder extends CustomAudioEncoder {
frame!: NodeAv.Frame;
packet!: NodeAv.Packet;
codecContext: NodeAv.CodecContext | null = null;
resampler: NodeAv.SoftwareResampleContext | null = null;
dstFrame: NodeAv.Frame | null = null;
avCodec!: NodeAv.Codec;
firstExpectedTimestamp: number | null = null;
outputTimestampOffset = 0;
inputParametersKey: string | null = null;
resamplerInputSampleRate: number | null = null;
nextResamplerPts: bigint | null = null;
packetEmitted = false;
adtsHeaderTemplate: AdtsHeaderTemplate | null = null;
static override supports(codec: AudioCodec, config: AudioEncoderConfig): boolean {
const { numberOfChannels, sampleRate } = config;
return (
codec === 'aac' && numberOfChannels >= 1 && numberOfChannels <= 48
&& AAC_SAMPLE_RATES.includes(sampleRate)
) || (
codec === 'opus' && numberOfChannels >= 1 && numberOfChannels <= 255
&& OPUS_SAMPLE_RATES.includes(sampleRate)
) || (
codec === 'mp3' && numberOfChannels >= 1 && numberOfChannels <= 2
&& MP3_SAMPLE_RATES.includes(sampleRate)
) || (
codec === 'vorbis' && numberOfChannels >= 1 && numberOfChannels <= 255
&& sampleRate <= 200000
) || (
codec === 'flac' && numberOfChannels >= 1 && numberOfChannels <= 8
&& sampleRate <= 655350
) || (
codec === 'ac3' && numberOfChannels >= 1 && numberOfChannels <= 6
&& AC3_SAMPLE_RATES.includes(sampleRate)
) || (
codec === 'eac3' && numberOfChannels >= 1 && numberOfChannels <= 16
&& AC3_SAMPLE_RATES.includes(sampleRate)
) || (
codec === 'dts' && DTS_CHANNEL_COUNTS.includes(numberOfChannels)
&& DTS_SAMPLE_RATES.includes(sampleRate)
&& dtsBitrateFits(resolveBitrate(config, codec), sampleRate, numberOfChannels)
);
}
async init(): Promise<void> {
this.frame = new NodeAv.Frame();
this.frame.alloc();
this.packet = new NodeAv.Packet();
this.packet.alloc();
const codecId = CODEC_TO_CODEC_ID[this.codec];
assert(codecId !== undefined);
const codec = NodeAv.Codec.findEncoder(codecId);
if (!codec) {
throw new Error(`Unable to obtain libav codec for '${this.codec}'.`);
}
this.avCodec = codec;
await this.createCodecContext();
}
async createCodecContext() {
assert(this.codecContext === null);
const codecContext = new NodeAv.CodecContext();
codecContext.allocContext3(this.avCodec);
let sampleFormat = NodeAv.AV_SAMPLE_FMT_FLTP;
if (this.avCodec.sampleFormats && !this.avCodec.sampleFormats.includes(NodeAv.AV_SAMPLE_FMT_FLTP)) {
// Use a format that's supported
sampleFormat = this.avCodec.sampleFormats[0]!;
}
codecContext.sampleRate = this.config.sampleRate;
codecContext.channelLayout = this.codec === 'dts'
? getDtsChannelLayout(this.config.numberOfChannels)
: getChannelLayout(this.config.numberOfChannels);
codecContext.codecType = NodeAv.AVMEDIA_TYPE_AUDIO;
codecContext.codecId = CODEC_TO_CODEC_ID[this.codec]!;
codecContext.sampleFormat = sampleFormat;
codecContext.timeBase = new NodeAv.Rational(1, this.config.sampleRate);
codecContext.bitRate = BigInt(resolveBitrate(this.config, this.codec));
if (this.config.bitrateMode === 'constant') {
codecContext.rcMinRate = codecContext.bitRate;
codecContext.rcMaxRate = codecContext.bitRate;
}
// libav's DTS encoder is marked experimental, so it refuses to open at the default compliance level
const options = this.codec === 'dts'
? NodeAv.Dictionary.fromObject({ strict: -2 })
: null;
const ret = await codecContext.open2(this.avCodec, options);
NodeAv.FFmpegError.throwIfError(ret, 'Open codec context');
this.codecContext = codecContext;
}
async encode(audioSample: AudioSample): Promise<void> {
if (this.codecContext === null) {
await this.createCodecContext();
assert(this.codecContext);
}
this.firstExpectedTimestamp ??= audioSample.timestamp;
if (audioSample._data instanceof AvFrameAudioSampleResource) {
// Release any buffers still referenced from the previous encode before reffing the new frame, otherwise
// av_frame_ref leaks them
// https://github.com/Vanilagy/mediabunny/issues/392
this.frame.unref();
this.frame.ref(audioSample._data.frame);
} else {
copyAudioSampleToAvFrame(audioSample, this.frame);
}
this.frame.pts = BigInt(Math.round(audioSample.timestamp * this.config.sampleRate));
this.frame.duration = BigInt(Math.round(audioSample.duration * this.config.sampleRate));
this.frame.timeBase = new NodeAv.Rational(1, this.config.sampleRate);
const key = `${this.frame.sampleRate}:${this.frame.channels}:${this.frame.format}`;
if (this.inputParametersKey !== null && this.inputParametersKey !== key) {
throw new Error(
'Input audio parameters changed. For this audio encoder, you cannot change the input audio'
+ ' parameters over time.',
);
}
this.inputParametersKey = key;
// We need the resampler when:
// 1. Format conversion is needed (sample format, sample rate, or channel count differs)
// 2. The codec requires fixed frame sizes
const requiresResampler
= this.codecContext.frameSize > 0
|| this.codecContext.sampleFormat !== this.frame.format
|| this.codecContext.sampleRate !== this.frame.sampleRate
|| this.codecContext.channels !== this.frame.channels;
if (requiresResampler) {
if (!this.resampler) {
this.resampler = new NodeAv.SoftwareResampleContext();
this.resamplerInputSampleRate = this.frame.sampleRate;
// Resample straight to whatever layout the encoder was opened with, since not every codec accepts
// the layout that getChannelLayout hands out for a given channel count
const outLayout = this.codecContext.channelLayout;
const inLayout = getChannelLayout(this.frame.channels);
const ret = this.resampler.allocSetOpts2(
outLayout, this.codecContext.sampleFormat, this.codecContext.sampleRate,
inLayout, this.frame.format as NodeAv.AVSampleFormat, this.frame.sampleRate,
);
NodeAv.FFmpegError.throwIfError(ret, 'allocSetOpts2');
const ret2 = this.resampler.init();
NodeAv.FFmpegError.throwIfError(ret2, 'init');
this.dstFrame = new NodeAv.Frame();
this.dstFrame.alloc();
this.dstFrame.channelLayout = outLayout;
this.dstFrame.sampleRate = this.codecContext.sampleRate;
this.dstFrame.format = this.codecContext.sampleFormat;
this.dstFrame.nbSamples = this.codecContext.frameSize || FRAME_SIZE_FALLBACK;
this.dstFrame.duration = BigInt(this.dstFrame.nbSamples);
this.dstFrame.allocBuffer();
this.nextResamplerPts = this.frame.pts;
}
const inputBuffers = this.frame.data;
if (!inputBuffers) {
throw new DOMException('Frame has no data', 'EncodingError');
}
await this.resampler.convert(null, 0, inputBuffers, this.frame.nbSamples);
await this.pullResampledFrames();
} else {
await this.sendFrameAndReceivePackets(this.frame);
}
}
async pullResampledFrames() {
assert(this.codecContext);
assert(this.resampler);
assert(this.dstFrame);
assert(this.nextResamplerPts !== null);
const frameSize = this.codecContext.frameSize || FRAME_SIZE_FALLBACK;
while (true) {
const available = this.resampler.getOutSamples(0);
if (available < frameSize) {
break;
}
await this.resampler.convert(this.dstFrame.data, frameSize, null, 0);
this.dstFrame.pts = this.nextResamplerPts;
await this.sendFrameAndReceivePackets(this.dstFrame);
this.nextResamplerPts += BigInt(frameSize);
}
}
async sendFrameAndReceivePackets(frame: NodeAv.Frame | null) {
assert(this.codecContext);
const ret = await this.codecContext.sendFrame(frame);
NodeAv.FFmpegError.throwIfError(ret, 'Send frame');
while (true) {
const receiveRet = await this.codecContext.receivePacket(this.packet);
if (receiveRet === NodeAv.AVERROR_EAGAIN || receiveRet === NodeAv.AVERROR_EOF) {
break;
}
this.receivePacket(receiveRet);
}
}
receivePacket(ret: number) {
assert(this.codecContext);
assert(this.firstExpectedTimestamp !== null);
NodeAv.FFmpegError.throwIfError(ret, 'Receive packet');
if (!this.packet.data) {
return;
}
let timestamp = Number(this.packet.pts) / this.codecContext.sampleRate;
const duration = Number(this.packet.duration) / this.codecContext.sampleRate;
let data: Uint8Array = this.packet.data;
let metadata: EncodedAudioChunkMetadata | undefined;
if (this.packetEmitted) {
metadata = {};
} else {
// To compensate for any negative timestamp things that FFmpeg might do. It does these for a reason, to
// indicate encoder delay, but the notion of this is not yet supported in Mediabunny. Yes, this technically
// introduces audio sync drift.
this.outputTimestampOffset = Math.max(this.firstExpectedTimestamp - timestamp, 0);
const codecString = this.config.codec;
let description = this.codecContext.extraData
? toUint8Array(this.codecContext.extraData)
: undefined;
if (this.codec === 'aac') {
if (!description) {
throw new Error('Extradata expected for AAC.');
}
// eslint-disable-next-line @stylistic/max-len
// eslint-disable-next-line @typescript-eslint/no-explicit-any, @typescript-eslint/no-unsafe-member-access
const isAdts = (this.config as any).aac?.format === 'adts';
if (isAdts) {
const parsedConfig = parseAacAudioSpecificConfig(description);
this.adtsHeaderTemplate = buildAdtsHeaderTemplate(parsedConfig);
description = undefined; // Not used with 'adts' format
}
} else if (this.codec === 'opus') {
if (!description) {
// Technically not required by the WebCodecs/Mediabunny Codec Registry, but we strive to be better
throw new Error('Extradata expected for Opus.');
}
} else if (this.codec === 'vorbis') {
if (!description) {
throw new Error('Extradata expected for Vorbis.');
}
} else if (this.codec === 'flac') {
if (!description) {
throw new Error('Extradata expected for FLAC.');
}
// FFmpeg uses the STREAMINFO block as the extradata, but WebCodecs wants a different format:
// 1. The bytes 0x66 0x4C 0x61 0x43 ("fLaC" in ASCII)
// 2. A metadata block (called the STREAMINFO block) as described in section 7 of [FLAC]
// 3. Other optional metadata blocks (not included here, because, well, they're optional)
description = new Uint8Array([
0x66, 0x4c, 0x61, 0x43, // 'fLaC'
128, 0, 0, description.byteLength,
...description,
]);
}
metadata = {
decoderConfig: {
codec: codecString,
sampleRate: this.codecContext.sampleRate,
numberOfChannels: this.codecContext.channels,
description,
},
};
}
if (this.adtsHeaderTemplate) {
const frameLength = data.byteLength + this.adtsHeaderTemplate.header.byteLength;
this.adtsHeaderTemplate.bitstream.pos = 30;
this.adtsHeaderTemplate.bitstream.writeBits(13, frameLength);
const final = new Uint8Array(this.adtsHeaderTemplate.header.byteLength + data.byteLength);
final.set(this.adtsHeaderTemplate.header, 0);
final.set(data, this.adtsHeaderTemplate.header.byteLength);
data = final;
}
timestamp += this.outputTimestampOffset;
const packet = new EncodedPacket(
data,
'key',
timestamp,
duration,
);
this.packetEmitted = true;
this.onPacket(packet, metadata);
}
async flush(): Promise<void> {
if (!this.codecContext) {
return;
}
outer:
if (this.resampler) {
assert(this.resamplerInputSampleRate !== null);
const currentOutSamples = this.resampler.getOutSamples(0);
if (currentOutSamples === 0) {
break outer; // Clean cut-off point
}
const frameSize = this.codecContext.frameSize || FRAME_SIZE_FALLBACK;
assert(currentOutSamples < frameSize); // Because if it's more, it would've already been retrieved
const inputSamplesNeeded = Math.ceil(
((frameSize - currentOutSamples) / this.codecContext.sampleRate) * this.resamplerInputSampleRate,
);
this.resampler.injectSilence(inputSamplesNeeded);
await this.pullResampledFrames();
}
await this.sendFrameAndReceivePackets(null);
this.codecContext.freeContext();
this.codecContext = null;
this.packetEmitted = false;
this.firstExpectedTimestamp = null;
this.outputTimestampOffset = 0;
this.adtsHeaderTemplate = null;
this.resampler?.free();
this.resampler = null;
this.inputParametersKey = null;
this.resamplerInputSampleRate = null;
this.nextResamplerPts = null;
this.dstFrame?.free();
this.dstFrame = null;
}
close(): MaybePromise<void> {
this.codecContext?.freeContext();
this.frame.free();
this.packet.free();
this.dstFrame?.free();
this.resampler?.free();
}
}
const resolveBitrate = (config: AudioEncoderConfig, codec: AudioCodec) => {
return config.bitrate ?? new Quality('medium')._toAudioBitrate(codec) ?? 0;
};
/**
* The DTS encoder needs each frame to be big enough to hold the per-channel side info and rejects the whole
* configuration when it isn't, so this mirrors the check from FFmpeg's dcaenc.c.
*/
const dtsBitrateFits = (bitrate: number, sampleRate: number, numberOfChannels: number) => {
if (bitrate < 32000 || bitrate > 3840000) {
return false;
}
const hasLfe = numberOfChannels === 6;
const fullbandChannels = numberOfChannels - (hasLfe ? 1 : 0);
const frameBits = 32 * Math.ceil(Math.ceil(bitrate * 512 / sampleRate) / 32);
const minFrameBits = 132 + (493 + 28 * 32) * fullbandChannels + (hasLfe ? 72 : 0);
return frameBits >= minFrameBits && frameBits <= 8 * DTS_MAX_FRAME_SIZE;
};