add experiment

[asr1617.git] / proposal.tex

diff --git a/proposal.tex b/proposal.tex
index a5125d0..a2134b5 100644 (file)
--- 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
 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
 
 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.
 
 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}
 
 \end{document}