CC 1.srt

1
00:00:00,149 --> 00:00:02,120
Hello, and welcome, dear viewer.

2
00:00:02,120 --> 00:00:07,270
If you are alive in the year 2021, chances
are you’ve used audio files before.

3
00:00:07,270 --> 00:00:09,669
No, not that kind, the other one.

4
00:00:09,669 --> 00:00:13,030
Think about the different digital audio formats
you know.

5
00:00:13,030 --> 00:00:17,439
There’s WAV, MP3, maybe you also know about
OGG or AAC.

6
00:00:17,439 --> 00:00:22,170
But I want to talk about FLAC, the free lossless
audio codec.

7
00:00:22,170 --> 00:00:24,330
But why should you care?

8
00:00:24,330 --> 00:00:29,070
Well, we’ll get to that later, but here
are my own reasons summarized.

9
00:00:29,070 --> 00:00:33,970
FLAC is an open standard, there’s an open-source
reference implementation, reasonably good

10
00:00:33,970 --> 00:00:37,730
documentation and it has medium implementing
difficulty

11
00:00:37,730 --> 00:00:42,680
There’s not much “easy” information
on FLAC out there

12
00:00:42,680 --> 00:00:47,730
Learning about FLAC gives insight into a lot
of aspects of digital audio, data compression,

13
00:00:47,730 --> 00:00:52,150
and data storage
I have spent at least 2 months rabbit-hole-ing

14
00:00:52,150 --> 00:00:58,250
myself into FLAC and implementing a decoder
from scratch for SerenityOS

15
00:00:58,250 --> 00:01:00,590
First some background on the codec itself.

16
00:01:00,590 --> 00:01:05,220
FLAC was released in 2001, which means it’s
one of the newest of the “first generation”

17
00:01:05,220 --> 00:01:09,150
of compressed audio codecs, eight entire years
later than MP3.

18
00:01:09,150 --> 00:01:13,840
It’s an open standard from beginning to
end and even discourages DRM measures.

19
00:01:13,840 --> 00:01:18,110
Although many of you may have never heard
of it, it holds an important position as being

20
00:01:18,110 --> 00:01:23,000
very fast to decode and therefore suitable
for “weaker” hardware such as embedded

21
00:01:23,000 --> 00:01:25,050
devices or even microcontrollers.

22
00:01:25,050 --> 00:01:30,640
Also, as you will see later, it’s an ideal
format for transcoding CD audio.

23
00:01:30,640 --> 00:01:32,530
So what’s so special about FLAC?

24
00:01:32,530 --> 00:01:35,060
It’s in the name: lossless.

25
00:01:35,060 --> 00:01:38,820
When we talk about file formats, generally
there are three kinds.

26
00:01:38,820 --> 00:01:44,270
The uncompressed are the oldest and simplest
because all they do is store data.

27
00:01:44,270 --> 00:01:50,021
That is your BMP for images, WAV or CDA for
audio, and – wait, there aren’t really

28
00:01:50,021 --> 00:01:54,200
uncompressed formats for video because file
sizes would be ridiculous.

29
00:01:54,200 --> 00:01:58,290
That’s the problem with this approach in
general, even though it’s easy to come up

30
00:01:58,290 --> 00:02:00,700
with such a standard and implement it.

31
00:02:00,700 --> 00:02:04,110
Next, we have compressed lossless formats.

32
00:02:04,110 --> 00:02:09,599
These store all the data that originally existed,
nothing more and nothing less, but take less

33
00:02:09,599 --> 00:02:12,150
space than the original data with clever tricks.

34
00:02:12,150 --> 00:02:17,799
You see, sensible data that we humans produce
and recognize isn’t really random, it has

35
00:02:17,799 --> 00:02:20,569
lots of patterns.

36
00:02:20,569 --> 00:02:26,220
Compression algorithms exploit these patterns
to remove information that can be fully reconstructed.

37
00:02:26,220 --> 00:02:29,099
For example, imagine a piece of English text.

38
00:02:29,099 --> 00:02:32,279
The word “the” would be pretty common,
right?

39
00:02:32,279 --> 00:02:37,190
So what a text compressing format can do is
instead of storing “the” a thousand times,

40
00:02:37,190 --> 00:02:42,299
it just stores: “well, the word “the”
should be at this position, that position,

41
00:02:42,299 --> 00:02:45,629
and these other 998 positions”.

42
00:02:45,629 --> 00:02:47,959
That will take up much less space.

43
00:02:47,959 --> 00:02:50,680
In fact: it’s pretty much what ZIP does.

44
00:02:50,680 --> 00:02:55,530
The same thing can be done with non-text data,
except other, even more clever patterns are

45
00:02:55,530 --> 00:02:56,530
exploited.

46
00:02:56,530 --> 00:03:02,219
This way, we get PNG for images, FLAC for
audio, and Google’s VP9 for video.

47
00:03:02,219 --> 00:03:05,680
You see, here is where FLAC belongs!

48
00:03:05,680 --> 00:03:09,739
But just to finish up, we of course can go
one step further.

49
00:03:09,739 --> 00:03:13,329
Lossy compression starts from the same point
as lossless compression does.

50
00:03:13,329 --> 00:03:18,439
Except this time, we do another trick: Many
high-information types of data, like audio,

51
00:03:18,439 --> 00:03:23,469
video, and images, can be stored at a level
of detail that humans can’t even distinguish.

52
00:03:23,469 --> 00:03:27,760
We can’t hear quiet sounds or small differences
in sound.

53
00:03:27,760 --> 00:03:32,010
We can’t see minor differences in color
or small fluctuations in light.

54
00:03:32,010 --> 00:03:36,199
We can distinguish even less detail in moving
pictures.

55
00:03:36,199 --> 00:03:41,699
So lossy compression just throws away all
the data that a human is unlikely to notice

56
00:03:41,699 --> 00:03:42,699
anyways.

57
00:03:42,699 --> 00:03:49,629
Here, we finally get JPEG for images, MP3
for audio, and H.264 for video, which is the

58
00:03:49,629 --> 00:03:52,579
thing you most likely call MP4.

59
00:03:52,579 --> 00:03:57,889
So FLAC sits in between the uncompressed codecs
and the lossy codecs, as being both compressed

60
00:03:57,889 --> 00:03:59,589
and lossless.

61
00:03:59,589 --> 00:04:04,010
I have been interested in FLAC for quite some
time now, and one time around the beginning

62
00:04:04,010 --> 00:04:12,559
of May, I decided to create a FLAC loader
and decoder for SerenityOS’s LibAudio.

63
00:04:12,559 --> 00:04:16,549
If you haven’t yet heard of the SerenityOS
project, it’s a from-scratch graphical Unix

64
00:04:16,549 --> 00:04:19,959
operating system that is visually based on
90’s user interfaces.

65
00:04:19,959 --> 00:04:23,909
It’s a three-year-old open-source project
in an early state, so there’s a bunch of

66
00:04:23,909 --> 00:04:25,050
work to be done.

67
00:04:25,050 --> 00:04:26,250
Like implementing audio formats.

68
00:04:26,250 --> 00:04:32,730
Anyways, if you want to try out SerenityOS
or even contribute, there are links everywhere.

69
00:04:32,730 --> 00:04:37,000
After having the struggle of my life for two
months, now I understand the FLAC standard

70
00:04:37,000 --> 00:04:39,620
well enough to make these videos.

71
00:04:39,620 --> 00:04:42,750
So that hopefully other people won’t struggle
as much.

72
00:04:42,750 --> 00:04:44,530
Let’s go.

73
00:04:44,530 --> 00:04:49,759
Before we start with anything that is specific
to FLAC, it’s good to have a baseline understanding

74
00:04:49,759 --> 00:04:51,560
of digital audio.

75
00:04:51,560 --> 00:04:54,810
To understand digital audio, we need to understand
audio itself.

76
00:04:54,810 --> 00:04:58,600
You’ve probably already heard this, but
fundamentally, sound is nothing more than

77
00:04:58,600 --> 00:05:03,669
rapidly changing air pressure, created by
natural objects such a string, a resonating

78
00:05:03,669 --> 00:05:08,970
body of air, or much simpler, some fancy stuff
in your body that lets you speak.

79
00:05:08,970 --> 00:05:13,990
When I say rapid changes, I mean between about
20 and 20 thousand changes per second.

80
00:05:13,990 --> 00:05:19,050
These vibrations happen in the pattern of
waves travelling through a medium like air,

81
00:05:19,050 --> 00:05:25,020
one wave is a high point and a low point in
pressure that follow after one another.

82
00:05:25,020 --> 00:05:30,159
How often such a wave is produced is called
its frequency, measured in Hertz.

83
00:05:30,159 --> 00:05:33,099
Higher frequency equals more waves per second.

84
00:05:33,099 --> 00:05:37,620
That’s what humans can hear, more specifically,
that’s what all of this complicated biology

85
00:05:37,620 --> 00:05:40,710
in your ear can decipher into high and low
pitches.

86
00:05:40,710 --> 00:05:45,419
But let’s suppose that you can’t afford
to have a private news speaker whenever you

87
00:05:45,419 --> 00:05:49,620
need them, or even an orchestra at your disposal
whenever you wanna listen to some music.

88
00:05:49,620 --> 00:05:51,729
I know, hard to imagine.

89
00:05:51,729 --> 00:05:56,289
For that, we need to record some audio so
that we can play it back later.

90
00:05:56,289 --> 00:06:01,479
Microphones and the likes can translate air
pressure to electricity, so now we’re in

91
00:06:01,479 --> 00:06:03,259
the electronic realm.

92
00:06:03,259 --> 00:06:08,610
But there are still two major problems with
this signal that’s now expressed in voltage.

93
00:06:08,610 --> 00:06:10,110
Can you guess?

94
00:06:10,110 --> 00:06:13,960
It’s continuous, in time and in gain.

95
00:06:13,960 --> 00:06:19,560
No matter how close you pick two points of
time on the signal, there’s always an infinite

96
00:06:19,560 --> 00:06:22,650
amount of information in between them.

97
00:06:22,650 --> 00:06:27,789
And additionally, the signal’s actual gain
can take any value, at least any value loud

98
00:06:27,789 --> 00:06:31,310
enough and also quiet enough that your microphone
could pick it up.

99
00:06:31,310 --> 00:06:36,590
That’s just what we call an analog electric
audio signal, it’s simply a direct translation

100
00:06:36,590 --> 00:06:40,810
of another analog signal, in this case, the
air pressure.

101
00:06:40,810 --> 00:06:45,550
But this becomes an issue once we want a computer
to talk to the audio world.

102
00:06:45,550 --> 00:06:49,860
The information here is way too much for us
to store in a computer, in fact, it’s an

103
00:06:49,860 --> 00:06:52,469
infinite amount of information!

104
00:06:52,469 --> 00:06:57,159
And as you know, all computers can store is
numbers, finitely many, and those numbers

105
00:06:57,159 --> 00:06:59,639
are also not continuous.

106
00:06:59,639 --> 00:07:04,289
So how do we take this analog signal and turn
it into something we might call digital, something

107
00:07:04,289 --> 00:07:08,560
that a computer can take in, process, store,
transmit, and do all the other fun stuff that

108
00:07:08,560 --> 00:07:10,449
we like to do with data.

109
00:07:10,449 --> 00:07:13,330
The key is that humans are really bad at hearing.

110
00:07:13,330 --> 00:07:17,620
For starters, I already mentioned that we
can’t hear changes in air pressure faster

111
00:07:17,620 --> 00:07:19,150
than a certain point.

112
00:07:19,150 --> 00:07:24,879
20 kHz is actually just an upper limit of
the highest frequency, most adults can’t

113
00:07:24,879 --> 00:07:26,379
hear that anymore.

114
00:07:26,379 --> 00:07:30,860
And second, below a certain threshold , we
can’t discern the loudness of two different

115
00:07:30,860 --> 00:07:34,030
sounds at slightly different volumes.

116
00:07:34,030 --> 00:07:38,979
So what we can do is preserve just enough
information so that a human can’t figure

117
00:07:38,979 --> 00:07:43,340
out the difference between a real and a “fake”
audio signal.

118
00:07:43,340 --> 00:07:49,080
The first step is to remove the time continuity
of the signal, making it discrete in time.

119
00:07:49,080 --> 00:07:54,439
We do that through a process called “sampling”:
At regular intervals, we take note of the

120
00:07:54,439 --> 00:08:00,430
strength of the signal, in our case the voltage,
removing all the information in between . I’ll

121
00:08:00,430 --> 00:08:04,680
not go into the details here, but because
of something called the Nyquist-Shannon sampling

122
00:08:04,680 --> 00:08:10,620
theorem, it is mathematically proven that
we can recreate any frequency that is sampled

123
00:08:10,620 --> 00:08:16,969
at least twice per oscillation , meaning that
if the highest necessary frequency has 20,000

124
00:08:16,969 --> 00:08:21,999
oscillations per second, we need to sample
it 40,000 times per second.

125
00:08:21,999 --> 00:08:26,300
The term “Hertz” for describing how often
something happens per second is also used

126
00:08:26,300 --> 00:08:32,099
for sampling, so for this, we would say a
“sampling rate” of 40 kHz.

127
00:08:32,099 --> 00:08:39,290
You might have heard of the sampling rates
44.1 kHz or 48 kHz.

128
00:08:39,290 --> 00:08:43,710
Those are the most common ones and the reason
that they’re higher than 40 kHz is unfortunately

129
00:08:43,710 --> 00:08:48,880
not something I can go into today .
And second, we need to remove the remaining

130
00:08:48,880 --> 00:08:51,500
value continuity of these sample points.

131
00:08:51,500 --> 00:08:56,820
That’s called quantization . We translate
these voltages into whole numbers , where

132
00:08:56,820 --> 00:09:01,590
the lowest number represents the lowest signal
strength and the highest number the highest

133
00:09:01,590 --> 00:09:02,700
signal strength.

134
00:09:02,700 --> 00:09:05,800
Now, how many numbers do we need here?

135
00:09:05,800 --> 00:09:09,990
The critical question is again; at what point
can a human no longer hear the difference

136
00:09:09,990 --> 00:09:11,440
in volume?

137
00:09:11,440 --> 00:09:16,710
If we had six different values, we could have
three different volumes for an audio wave.

138
00:09:16,710 --> 00:09:20,320
That’s almost a difference between 30% and
70% volume, and you can clearly hear that.

139
00:09:20,320 --> 00:09:25,680
But it turns out that once we reach about
50,000 different volumes for signals, the

140
00:09:25,680 --> 00:09:28,600
human hearing stops noticing a difference.

141
00:09:28,600 --> 00:09:33,260
We call whatever amount of different values
we allow our “bit depth”.

142
00:09:33,260 --> 00:09:36,900
Because in reality these numbers will all
be stored in binary by the computer, so it

143
00:09:36,900 --> 00:09:40,550
makes sense that the range in values is a
power of two.

144
00:09:40,550 --> 00:09:46,610
By the bit depth, we don’t directly mean
the number of possible values, like 65,536,

145
00:09:46,610 --> 00:09:52,280
we instead mean the number of bits we use
to store one such sample value, like 16.

146
00:09:52,280 --> 00:09:58,510
As I’ve already said, a bit depth of 16
is about at the limit of what humans can hear,

147
00:09:58,510 --> 00:10:04,650
so bit depths below that are uncommon, but
above it are sometimes seen.

148
00:10:04,650 --> 00:10:06,530
So let’s revisit that.

149
00:10:06,530 --> 00:10:11,320
What we have now is a list of numbers, each
representing the strength of an audio signal

150
00:10:11,320 --> 00:10:17,070
at some point in time, with all these points
in time regularly spaced according to a sample

151
00:10:17,070 --> 00:10:18,190
rate.

152
00:10:18,190 --> 00:10:23,600
The bit rate of the numbers tells us the maximum
and minimum values, which represent the real-world

153
00:10:23,600 --> 00:10:26,950
maximum and minimum pressure or voltage.

154
00:10:26,950 --> 00:10:32,840
This is called pulse code modulation, PCM,
and it’s by far the most important way of

155
00:10:32,840 --> 00:10:35,630
digitizing any signal, not just audio.

156
00:10:35,630 --> 00:10:37,630
There are variations we can do here [non-linear
quantization, dynamic sample rates], but remember

157
00:10:37,630 --> 00:10:44,840
that this is all because circuits, real digital-analog
electronic hardware components, can do these

158
00:10:44,840 --> 00:10:51,950
conversions super easily and accurately, both
going from analog to digital, and from digital

159
00:10:51,950 --> 00:10:54,710
to analog.

160
00:10:54,710 --> 00:10:59,130
And unfortunately for you, the early viewer,
this is where I’ll end the first video of

161
00:10:59,130 --> 00:11:00,170
the FLAC series.

162
00:11:00,170 --> 00:11:05,050
I’m planning to cover every single part
of the specification, including, what you

163
00:11:05,050 --> 00:11:08,400
saw today, all the audio fundamentals you’ll
have to know.

164
00:11:08,400 --> 00:11:13,330
In episode 2, we’ll have a look at how we
can losslessly compress the audio data we

165
00:11:13,330 --> 00:11:19,220
have now obtained, and in episode 3, we’ll
peek inside how FLAC stores all of this stuff.