Hi, my name is Randall Chu and in this video I will be presenting a summary of my final
year project: Computerised Chord Transcription in Simple Electronic Dance Music.
I will start by giving a high level summary of my project.
After that I will enter into a discussion that delves more deeply into the details of
my work.
To conclude, we will have the chance to hear some music and see firsthand how the transcriptions
perform.
Let us begin.
In music, chords are combinations of notes produced by instruments.
They vary over time, typically changing every few seconds.
Progressions of chords, are what create mood and feel in songs.
Musicians need to know chords in order to play, and often label these by hand.
However, this is very time-consuming.
Thus, there is interest in automating this process.
Work has been done in this area over the last few decades.
However, it has tended to focus on genres such as pop, rock and classical.
In this project, I consider a different genre, electronic dance music.
This genre is interesting from a transcription perspective because the chords tend to be
simple, but there are also noisy features which obscure the chords.
You may not have seen chords before, so here is a demonstration.
Simply pay attention to the bottom row, and you should be able to perceive that something
in the background is changing when the chord label changes.
Now that you've seen what chord labels look like, you should be able to appreciate that
the goal of my project was to perform the transcription task using a computer.
Specifically, my aim was to design and implement a method that was able to handle the unique
features of electronic dance music, including segments with no chords, and noisy instrumentation.
I also wanted to achieve an accuracy that was on par with existing methods.
To narrow the scope to what is appropriate for a year-long project, I made simplifying
assumptions about the chord content in the music.
This is why my project title identifies Simple Electronic Dance Music.
To approach this problem, I modelled chord transitions with a Hidden Markov Model, using
chromagrams as observations.
I used supervised learning, with a set of songs that I transcribed myself as the data
set.
In the transcriptions themselves, there was good success with drops and moderate success
with heavy percussion, sweeps and the no-chord scenario.
I found that giving preferential treatment to the mid-ranged frequencies when generating
the chromagram gave improved results.
When doing this, the average transcription accuracy was 70%; slightly below the 80-90%
that state-of-the-art algorithms achieve elsewhere.
That concludes the high level summary.
We will now explore the project in greater detail.
A good place to start is a precise statement of the problem we are trying to solve.
Given an input audio file, we wish to determine both the chords and the boundaries of the
chords in time.
In this project I have assumed that the audio contains only major and minor chords, for
simplicity.
This means there are 24 possible chords to choose from.
I have also assumed that the chords are built off the major scale.
This means that if we can identify what is called the "key" of a song, we only need
to choose from 6 out of the 24 chords at any one time.
As you might have seen in the demonstration earlier, the transcriptions generated assume
that the key is known, because they used the roman numerals.
This is a reasonable assumption, if we have a good key detection system available.
However, in electronic dance music, we will not have just the 6 chords, we will also have
an additional state to accommodate the scenario where there is `no-chord'.
Thus there are 7 states in total to consider.
To estimate the states, I first divided up the music into 300 ms segments, with the aim
of finding the most likely chord in each segment.
I then the computed the spectrogram for each segment, which provided the distribution of
different frequencies in the audio.
These were then mapped to the frequencies corresponding to each of the different notes
in music.
There are 12 notes, and each chord being transcribed contains 3 of the 12 notes.
Since not all frequency content contributes equally to the chord, one of my strategies
to improve the chromagram was to give less weight to certain frequencies, when summing
the 88-bin chromagram to create the 12-bin chromagram.
The 88 bins simply reflect the 88 keys on a piano, which is representative of the range
of audible frequencies.
Hence, for each 300 ms segment, we have a chord that is not directly visible, but frequency
information that is visible in the form of a chromagram.
We also know that these chords change over time.
We can model this using a Hidden Markov Model, which assumes that the observations have some
distribution, and that the state transition probabilities are fixed.
To train the Hidden Markov Model, I used a set of songs, the training set, labelled with
the chords, and modelled the chromagrams for each chord state using a Gaussian distribution.
After estimating the distributions and the state transition matrix were estimated, I
then took another set of unseen songs, the test set, and tried to decode the chords using
only the chromagrams.
The tool for doing this decoding is the Viterbi Algorithm.
So, using this setup how accurate were the transcriptions?
To measure them, what I did, was take the parameters estimated on the training set,
and ran the Viterbi Algorithm on both training and test sets, assuming that the chord sequence
was hidden.
As expected, the test set score is slightly below the training set because this represents
a new, previously unseen set of songs.
We see that the accuracy averages 56% using unweighted chromagrams.
If we use weighted chromagrams, we get an improvement of nearly 15%, to an accuracy
of 70%.
I also undertook to randomise the training and test data sets and ran the model to get
an average over 1,000 allocations.
Using the weighted chromagrams, the average accuracy was still at the 70% mark, plus or
minus 5%.
There is also the question of how well the transcriptions did with the unique features
of electronic dance music.
To summarise, they did fairly well with the no-chord state at a feature called a "drop",
which is when the instrumentation decreases suddenly to create anticipation.
They also did moderately well in the presence of heavy percussion.
It was interesting to note that uncertain and ambiguous audio segments tended to be
labelled with a 'no-chord', rather than an incorrect chord.
We'll get to see many examples of these in a moment.
To summarise, in this project I investigated electronic dance music with simple chord progressions,
and modelled chords using a Hidden Markov Model.
The key innovations in this project were the use of weighted chromagrams, and the inclusion
of an additional state for the 'no-chord' situation.
It was observed that the weighted chromagram improved the transcription accuracy, and that
the no-chord state lead to successful transcriptions, with some interesting behaviour in noisy audio.
This project has shown that there is warrant and feasibility to transcribing the no-chord
state, which is currently not emphasised in the literature.
Hidden Markov Models are comparatively simple compared to other transcription methods, and
are no longer considered state-of-the-art, yet they have worked surprisingly well in
simple electronic dance music.
That brings us to the end of the detailed overview of my project.
We now have the chance to see how the HMM performs.
So sit back, enjoy the music, and pay attention to where the transcriptions fail and see if
you can identify anything interesting.
That brings us to the end of this presentation.
I hope that you have a better appreciation of the role of chords in music, and some of
the challenges in automated transcription.
Thank you for your time.
For more infomation >> 【中華먹방】コチュチャプチェが好きすぎて。。 - Duration: 9:48. 
For more infomation >> anekdoty pro brak - Duration: 0:41. 
For more infomation >> Мультики про машинки с игрушками Плеймобил - Выносливость и полиция! Игрушечные видео для мальчиков - Duration: 5:48. 
For more infomation >> 塔罗占卜：最近你的工作会有提升吗？ - Duration: 2:17. 
For more infomation >> Landesschau Baden-Württemberg vom 18.10.2018 - Duration: 44:47. 
For more infomation >> Тест! ЧТО ВЫ УВИДЕЛИ ПЕРВЫМ на картинке. Ваша Судьба в Одном Изображении. Тест Точен на 98%! - Duration: 1:46. 
For more infomation >> Wie Frauen und Männer Junggesellenabschiede feiern | Landesschau Baden-Württemberg - Duration: 4:10. 
For more infomation >> Brandmelder halten Feuerwehr auf Trab | Landesschau Baden-Württemberg - Duration: 4:30. 
For more infomation >> ПРИВИВКА ОТ ГРИППА. Вакцинация от гриппа. Делать или нет? Вред и последствия! - Duration: 2:07. 
For more infomation >> Rekord-Niedrigwasser am Rhein | Landesschau Baden-Württemberg - Duration: 2:35. 
For more infomation >> Zurück vom Hilfseinsatz in Indonesien | Landesschau Baden-Württemberg - Duration: 7:41. 
Không có nhận xét nào:
Đăng nhận xét