From c27924e430557e2da1afc248eada4bd0c08f50aa Mon Sep 17 00:00:00 2001
From: Mart Lubbers <mart@martlubbers.net>
Date: Thu, 9 Mar 2017 19:11:55 +0100
Subject: [PATCH] finalize proposal
---
proposal.pre | 1 +
proposal.tex | 44 ++++++++++++++++++++++++++++++++++++++++++--
2 files changed, 43 insertions(+), 2 deletions(-)
diff --git a/proposal.pre b/proposal.pre
index 5f67969..c9cfd70 100644
--- a/proposal.pre
+++ b/proposal.pre
@@ -4,6 +4,7 @@
\usepackage{geometry} % Papersize
\usepackage{hyperref} % Hyperlinks
+\usepackage{booktabs} % Nice tables
\urlstyle{same}
\hypersetup{%
diff --git a/proposal.tex b/proposal.tex
index a5125d0..a2134b5 100644
--- a/proposal.tex
+++ b/proposal.tex
@@ -12,7 +12,9 @@ tackled using myriads of different approaches such as HMMs with different
acoustic model types but also machine learned feature sets. I would like to
explore how HMM based techniques perform on extreme heavy metal styles to see
how well it can detect growling and how classifier might be adapted to perform
-better.
+better. Initially the classifier will be a binary classifier that classifies
+growling and non-growling. Later on classes might be added such as and musical
+genres.
Singing voice detection is often used as a preprocessing step for song lyrics
forced alignment. If the time permits I would like to explore forced alignment
@@ -20,6 +22,44 @@ using existing phone models on extreme heavy metal styles. Features probably
need to be changed to improve performance since growling is very different from
regular singing and speaking.
-The data for this will be coming from my personal collection audio CDs.
+The data for this will be coming from my personal collection audio CDs. For the
+singing voice detection the data of one band can be used and even a test set
+can be held out since the band made over 15 full studio albums each with a
+running time between 30 and 60 minutes. The segmentation for the trainingsdata
+can be done by hand. Later, for the lyrics alignment, the labels for the
+segments can be found online on song lyric websites. Validation of the
+alignment is a bit tricky however since there is no golden standard but my own.
+
+\paragraph{Planning}\strut\\
+This results in the following rough outline divided on a month by month basis
+shown in Table~\ref{tbl:outline}.
+Possible pitfalls can arise in preparing the data since that requires
+segmentation. It is expected to take around twice the playing time but that
+might be an overestimation.
+
+\begin{table}[ht]
+ \centering
+ \begin{tabular}{cll}
+ \toprule
+ Month & Description\\
+ \midrule
+ March
+ & Preparing the data\\
+ & Preparing an experiment platform\\
+ & Literature research\\
+ April
+ & Running the experiments\\
+ & Fiddle with parameters\\
+ & Explore the possibilities for forced alignment\\
+ May
+ & Write up the thesis\\
+ & Possibly do forced alignment\\
+ June
+ & Finish up thesis\\
+ & Wrap up\\
+ \bottomrule
+ \end{tabular}
+ \caption{Outline}\label{tbl:outline}
+\end{table}
\end{document}
--
2.20.1