Just some stuff about me.
Here's my dotfiles repository.
Hey fellow penguins,
A few days ago, an user asked about audio quality on Linux, and whether it is worse or better than audio on Windows. The thread became a mess quickly, full of misconceptions and urban myths about Linux. I figured it would be worthwhile to create a complete guide to Linux audio, as well as dispelling some myths and misconceptions.
To all be on the same page, this is going to be thorough, slowly introducing more concepts.
You might remember from high school that sound is waves traveling through the air. Vibrations of any kind cause molecules in the air to move. When that wave form finds your ears, it causes little hairs in your ear to move. Different hairs are susceptible to different frequencies, and the signals sent by these hairs are turned into sound you hear by your brain.
In reality it is a little more complicated, but for the sake of this post, that’s all you need to know.
The pitch of sound comes from its frequency, the ‘shorter’ the waves are in a waveform, the higher the sound. The volume of sound comes from how ‘tall’ the waves are. Human hearing sits in a range between 20Hz and 20,000 Hz, though it deviates per person. Being the egocentric species we are, waves below 20 Hz are called ‘infrasound’ and waves above 20kHz are called ‘ultrasound.’ Almost no humans can hear beyond ultrasound, you will find that your hearing probably cuts off at 16kHz.
To play around with this, check out this tone generator, you can prove anything above with this yourself. As a fun fact: human hearing is actually really bad, we’ve among the most limited frequency ranges. A cat can hear up to 40kHz, and dolphins can even hear up to 160kHz!!
FACT: Playing loud music is dangerous! If you listen to music and you are feeling a discomfort, you should turn the volume down. A true alert is when you hear a beep - this is called tinnitus, and that beep you’re hearing is pretty much the death cry of the cells that can hear that frequency. That beep is the last time you will hear that very specific frequency ever again. Please, listening to loud music is not worth the permanent hearing damage, please dial it down for your own sake! <3
To listen to sound, you will probably be using headphones or speakers, inside of them are cones that are driven by an electromagnet, causing them to vibrate at very precise frequencies. This is essentially how sound works, though modern headphones certainly can be pretty complex.
To drive that magnet, an audio source will send an analog signal (a waveform) over a wire to the driver, causing it to move at the frequency of that waveform. This is in essence how audio playback works; and we’re not going to get into it much deeper than this.
Computers are digital - which is to say, they don’t do analog; processors understand ON and OFF, they do not understand 38.689138% OFF and 78.21156% ON. When converting an analog signal (like sound) to a digital one, we make use of a format called PCM. For PCM to be turned into an analog signal, you need a DAC - or as you probably know it: a sound card. DAC stands for ‘Digital to Analog Converter’, or some people mistakenly call it “Digital Audio Converter/Chip”
PCM stands for Pulse-code Modulation, which is a way to represent sampled analog signals in a digital format. We’re not going to get into it too much here, but imagine taking a sample of a waveform at regular intervals and storing the value, and then rounding that value to a nearest ‘step’ (remember this). That’s PCM.
The fidelity of PCM comes from two elements, which we are going to discuss next: sampling rate and bit depth.
Sampling rate is the most important part of making PCM sound good. Remember how humans hear in a range of 20Hz to 20kHz? The sample rate of audio has a lot to do with this. You cannot capture high frequencies if you do not capture samples often enough. Since our ears can hear up to 20 kHz, you would imagine that 20kHz would be ideal for capturing audio; however, a result of sampling is that you actually need twice the sample rate, this is called the Nyquist-Shannon sampling theorem, which is a complicated thing. Just understand that to reproduce a 20kHz frequency, you need a sample rate of 40kHz.
To have a little bit of room and leeway, we settled on a sample rate of 48kHz (a multiple of 8) for playback, and 96kHz for recording. We record at this frequency only to make sure absolutely no data is lost. You might be more familiar with 44.1kHz for audio, which is a standard we settled on for CD playback and NTSC. A lot of scientific research has been done on sound quality, and there is no evidence to suggest people can tell the difference between 48kHz or anything higher.
MYTH BUST: Humans cannot hear beyond 20 kHz, period. Anyone who claims to be able to is either supernatural or lying to you - I’ll let you choose which.
Remember how I told you to remember that PCM rounds values to the nearest step? This has to do with how binary works. The more bits, the bigger the number you can store. In PCM, the bit-depth decides the number of bits of information in each sample. With 16-bit, the range of values that can be stored is 0 to 65535. Going beyond this is pointless for humans, with no scientific research showing any proven benefit, though marketeers would like you to believe there’s benefits.
MYTH BUST: 24-bit depth is often touted as ‘high-resolution audio’, claiming benefits of a better sonic experience. Such is nothing more than marketing speech, there is no meaningful data 24-bit can capture that 16-bit cannot.
We’ll briefly touch on the last part of PCM audio, channels. This is very self explanatory, humans have two ears and can hear separate sounds on both of them, which means we have stereo hearing. As a result, most music is recorded with 2 channels. For some surround settings, you need more channels, this is why you may have heard of 5.1 or 7.1; the first digit is the amount of channels the PCM carries.
For most desktop usage, the only sound we care about is 2-channel PCM.
So, we’ve covered all the elements of PCM sound. Let’s go over it quickly: sample rate is expressed in Hz and is how often a sample of a waveform is captured, representing the x-axis of a waveform. Bit-depth is the bits of information stored in each sample, and represents the y-axis of the waveform. Channels decide how many simultaneous outputs the PCM can drive separately, since we have 2 ears, you need at least two channels.
As a result, the standard audio playback for both consumers and professionals is 48kHz, 16-bit, 2 channel PCM. This is more than enough to fully represent the full range of human hearing.
So, now that we know how PCM works, how does Linux make sound? How can you make Linux sound great? A few important components come into play here, and we’ll need to discuss each of them in some detail.
ALSA is the interface to the kernel’s sound driver. ALSA can take a PCM signal and send it to your hardware by talking to the driver. Something important to know about most DACs is that they can only take one signal at a time, actually. That means that only a single application can send sound to ALSA at once. Long ago, in a darker time, you couldn’t watch a movie while listening to music!
This problem was solved a long time ago with the use of alsalib, but doing mixing at a library level isn’t a very good solution to the problem. This gave rise to sound servers, of which many have existed. Before PulseAudio, esound was a very popular one but had many problems, eventually it was succeeded by PulseAudio.
When you think audio on Linux, PulseAudio is probably among the first things you think of. PulseAudio is NOT a driver, nor does it talk to your drivers. Actually, PulseAudio only does two things that we’ll discuss in detail later. PulseAudio talks to ALSA, taking control of its single audio stream, and allows other applications to talk to PulseAudio instead. Pulse is an ‘audio multiplexer’, turning multiple signals into one through a process that is called mixing. Mixing is an incredibly complicated subject that we won’t talk about here.
To be able to mix sounds, one must make sure that all the PCM sources are in the same format (the one that’s being sent to ALSA); if the PCM format being sent to Pulse does not match the PCM format being sent to ALSA, pulse does a step before mixing it called resampling. Resampling is another very complicated subject that can turn a 8kHz, 4-bit, 1-channel PCM stream into a 24kHz, 24-bit, 2-channel PCM stream.
These two things allow you to play a game, listen to music and watch YouTube, and notifications to produce a sound all at the same time. PulseAudio is the most critical element of the Linux sound stack.
FACT: PulseAudio is a contentious subject, many people have a dislike for this particular bit of software. In all honesty, PulseAudio was brought to the general public in a bit of a premature state, breaking audio for many people. PulseAudio these days is a very stable, solid piece of software. If you have audio issues these days, it’s usually a problem in ALSA or your driver.
PulseAudio isn’t the only sound server/daemon available for Linux, though it is certainly the most popular and most likely the default of whatever distribution you are using. PulseAudio has become a bit of a standard for Linux sound and has by far the best compatibility with most applications, but that doesn’t mean there aren’t alternatives.
JACK (JACK Audio Connection Kit, a recursive acronym like GNU) is a sound server focused primarily on low latency. If you are doing professional audio work on Linux, you will already be very familiar with JACK. JACK’s development is very focused on low latency, real-time audio and is critical for such people. JACK is available on most distros as an alternative, and you can try it for yourself if you so want; but you might find some applications do not work nicely with JACK.
PipeWire is a project that is currently in development, looking to solve key problems that exist in current sound servers. PipeWire isn’t just a sound server but also handles the multiplexing of video sources (like a camera). Special attention has been put into working with sandboxed applications (like Flatpaks), which is an area where PulseAudio is lacking. PipeWire is a very promising project that might very well succeed PulseAudio in the future and you should expect to see appearing in distribution repositories very soon. You can try it yourself right now, though it isn’t quite as easy to get started with as JACK is.
More audio servers exist, but are beyond the scope of this post.
Resampling is the process of turning a PCM stream into another PCM stream of a different resolution. Your DAC only accepts a limited range of PCM signals, and it is up to the software to make sure the PCM stream is compatible. There is almost no DAC out there that doesn’t support 44.1kHz, 16-bit, 2-channel PCM, so this tends to be the default. When you play an audio source (like an OggVorbis file), the PCM stream might be 96kHz, 24-bit, 2-channel PCM.
To fix that, PulseAudio will use a resampling algorithm. There are two kinds of resampling methods: upsampling and downsampling. Upsampling is lossless, since you can always represent less data with more data. Downsampling is lossy by definition, you cannot represent 24-bit PCM with 16-bit PCM.
MYTH: Downsampling is a loss in quality! This is only true in a technical sense, or if you are downsampling to less than 48kHz, 16-bit PCM. When you downsample a 96kHz, 24-bit PCM stream to a 48kHz, 16-bit stream, no meaningful data is lost in the process; because the discarded data lies outside of the human ear’s hearing range.
FACT: Resampling is expensive. Good quality resampling algorithms actually take a non-trivial amount of processing power. PulseAudio defaults to a resampling method with a good balance between CPU time used and quality.
Mixing is the process of taking two PCM streams and combining them into one. This is extremely complicated and not something we’re going to discuss at length. It is not important to understand how this works, only to understand that it exists. Without mixing, you wouldn’t be able to hear sounds from multiple sources. This is true not just for PulseAudio and computer sound, this is true for anything. In real life, you might use an A/V receiver to accept sound from your TV and music player at once, the receiver then mixes the signals and plays it through your speakers.
Finally we can talk a little about encoding. Encoding is the process of taking a PCM stream and writing it to a permanent format, two types exist. You have lossy encoding and lossless encoding. Lossy encoding removes data from the PCM stream to safe space. Usually the discarded data is useless to you, and will not make a difference in sound quality; examples of lossy encoding are MP3, AAC and Ogg Vorbis. Lossless encoding takes a PCM stream and encodes it in such a way that no data is lost, examples of lossless encodings are FLAC, ALAC and WAV.
Note that lossy and lossless do not mean compressed and uncompressed. A lossless format can be compressed and usually is, as uncompressed lossless encoding would be very large; it would just be the raw PCM stream. An example of lossless uncompressed audio is WAV.
A new element encodings bring is their bit rate, not to be confused with sample rate and bit depth. Bit rate has to do with how much data is stored in every second of audio. For a lossless, uncompressed PCM stream this is easy to calculate with the formula bit rate = sample rate * bit depth * channels, for 16-bit, 48kHz, 2 channel PCM this is 1,5 Mbit. To get the value in bytes, divide by 8, thus 192kB per second.
The bit rate of an encoder means how much the audio will be compressed. PCM compression is super complicated, but it generally involves discarding silence, cutting off frequencies you cannot hear, and so forth. Radio encoding has a bit rate of roughly 128 Kbps, while most CDs have a bit rate of 1360kbps.
Lastly, there is the concept of VBR and CBR. VBR stands for Variable Bit Rate, which CBR stands for Constant Bit Rate. In a VBR encoding, the encoder aim for a target bit rate that you set, but it can deviate if it thinks it needs more or less. CBR will encode a constant bit rate, and will never deviate.
MYTH: Lossless sounds better than lossy. This is blatantly untrue, lossless audio formats were created for preservation and archival reasons. When you encode a lossy file from a lossless source, and you make sure that it’s a 48kHz, 16-bit PCM encoding, you will not lose any important information. What is enough depends on the quality of the encoder. For OggVorbis, 192kbps is sufficient, for MP3, 256kbps should be preferred. 320kbps is excessive and the highest quality supported by MP3. In general, 256kbps does the trick, but with storage being abundant these days, you can play it safe and use 320kbps if it makes you feel better.
MYTH: CBR is better than VBR. There is no reason not to use VBR at all, there is no point in writing 256Kbps of data if there is only silence or a constant tone. Let your encoder do what it does best!
FACT: Encoding a lossy format to another lossy format will result in a loss of data! You will compress data that is already compressed, which is really bad. When encoding to a lossy format, always use a high quality recording in a lossless format as the source!
I DON’T BELIEVE YOU: This article from the guys Xiph (the people who brought you FLAC and Ogg Vorbis) explain it better than I can: https://people.xiph.org/%7Exiphmont/demo/neil-young.html
Here is a quick guide to achieving great sound quality on Linux with the above in mind.
/etc/pulse/daemon.conf file.
default-sample-* settings to your sound card.avoid-resampling to true, this is a huge misconception, this does not improve sound quality at best, and in the worst case, can actually break things.As you can see, there’s little you can do in Linux in the first place, so what can you do if you want better sound?
MYTH: Linux sound quality is worse than Windows. They are exactly the same, Pulse doesn’t work that different from how Windows does mixing and resampling.
MYTH: Linux sound quality can be better than Windows. They are exactly the same. All improvements in quality come from the driver and your DAC, not the sound server. Pulse and ALSA do not touch the PCM beyond moving it around and resampling it.
I hope this (long) guide was of help to you, and helped to dispel some myths. Did I miss anything? Ask or let me know, and I’ll answer the best I can. Did I make any factual errors? Please correct me with a source and I’ll amend the post immediately.