155 lines
7.1 KiB
TeX
155 lines
7.1 KiB
TeX
\documentclass{scrartcl}
|
|
\usepackage{enumitem}
|
|
\usepackage[british]{babel}
|
|
\usepackage[style=apa, backend=biber]{biblatex}
|
|
\DeclareLanguageMapping{british}{british-apa}
|
|
\usepackage{url}
|
|
\usepackage{float}
|
|
\restylefloat{table}
|
|
\usepackage{perpage}
|
|
\MakePerPage{footnote}
|
|
\usepackage{graphicx}
|
|
% Create hyperlinks in bibliography
|
|
\usepackage{hyperref}
|
|
|
|
\newsavebox{\abstractbox}
|
|
|
|
\usepackage{etoolbox}
|
|
\makeatletter
|
|
\expandafter\patchcmd\csname\string\maketitle\endcsname
|
|
{\vskip\z@\@plus3fill}
|
|
{\vskip\z@\@plus2fill\box\abstractbox\vskip\z@\@plus1fill}
|
|
{}{}
|
|
\makeatother
|
|
|
|
\DeclareCiteCommand{\citeyearpar}
|
|
{}
|
|
{\mkbibparens{\bibhyperref{\printdate}}}
|
|
{\multicitedelim}
|
|
{}
|
|
\usepackage[T1]{fontenc}
|
|
\usepackage[utf8]{inputenc}
|
|
\usepackage{blindtext}
|
|
\setkomafont{disposition}{\normalfont\bfseries}
|
|
|
|
\addbibresource{~/PerryPerrySource/LaTeX/Hud_masters.bib}
|
|
|
|
\begin{document}
|
|
|
|
\title{Huddersfield Research Masters}
|
|
\subtitle{Combined F0 Estimation Algorithm Proposal}
|
|
\author{Sam Perry}
|
|
\date{}
|
|
\maketitle
|
|
|
|
\begin{abstract}
|
|
The pitch of audio is a perceptually important characteristic as it
|
|
forms the building block for musical characteristics such as key,
|
|
melody, and harmony. Many methods have been developed for estimating
|
|
the fundamental frequency of a signal, however very few have come close
|
|
to generating estimations that resemble human perception of pitch with
|
|
the same level of detail. Factors such as noise and the absence of a
|
|
clear fundamental frequency cause erroneous results in algorithms and
|
|
the concept of polyphony further complicates the problem as this
|
|
requires the separation of different notes. The variety of estimation
|
|
algorithms available has lead to a selection of methods that each
|
|
perform to varying standards depending on conditions. For example,
|
|
time-domain approaches, such as the autocorrelation approach, are able
|
|
to detect the correct pitch of a signal more accurately than frequency
|
|
domain approaches, such as the harmonic-product spectrum method, when
|
|
the fundamental frequency is missing. However neither is able to detect
|
|
multiple pitches in the way that the MUSIC algorithm can.
|
|
\end{abstract}
|
|
|
|
\section{Overview}
|
|
This project would aim to explore the possibility of combining pre-existing
|
|
algorithms based on audio descriptor analyses in order to adaptively select
|
|
the algorithm with the best chance of an accurate estimate. The aim would
|
|
be to create a robust tool for offline (and potentially realtime)
|
|
estimation of F0 values.
|
|
|
|
\section{Background/Existing Techniques}
|
|
\subsection{F0 Estimation Techniques}
|
|
The most popular F0 estimation algorithms can be categorized as one of
|
|
two types:
|
|
\begin{itemize}
|
|
\item Spectral Techniques
|
|
\item Temporal Techniques
|
|
\end{itemize}
|
|
\subsubsection{Spectral Techniques}
|
|
Spectral techniques focus on analysing the spectral content of the signal
|
|
by using output from an FFT to perform further processing to determine the
|
|
F0 value. Types of technique that can be categorized in this way include:
|
|
\begin{itemize}
|
|
\item Harmonic Product Spectrum~\parencite[p.8]{smyth2015hps}
|
|
\item Cepstral analysis
|
|
\item Maximum likelihood
|
|
\end{itemize}
|
|
\subsubsection{Temporal Techniques}
|
|
Temporal techniques attempt to calculate the periodicity of the signal.
|
|
This can then be inverted to produce the frequency.
|
|
Types of technique that can be categorized in this way include:
|
|
\begin{itemize}
|
|
\item Autocorrelation~\parencite[p.98]{lerch2012itaca}
|
|
\item Zero-crossing~\parencite[p.98]{lerch2012itaca}
|
|
\item NCCF (normalized cross correlation function)~\parencite{kasi2015yaapt}
|
|
\end{itemize}
|
|
\subsection{Current methods for technique combination}
|
|
\label{sec:ComMeth}
|
|
There is considerably less research into the combination of multiple
|
|
algorithms for the improvement of results. Limited research has been
|
|
carried out into the effects of training supervised learning algorithms to
|
|
pick results based on circumstances.
|
|
The ``Yet Another Algorithm for Pitch Tracking'' algorithm attempts to
|
|
refine temporal analysis results through the analysis of spectral
|
|
information.~\cite{kasi2015yaapt}
|
|
Overall there remains a large scope for the type of research proposed.
|
|
|
|
\section{Methodology}
|
|
A detailed analysis of a variety of the most prominent F0 estimation
|
|
techniques will be presented, to determine the quantity of methods
|
|
needed and which methods will produce the best quality results. Methods
|
|
for selecting an algorithm will also require significant further
|
|
research. Potential techniques to be explored include:
|
|
\begin{itemize}
|
|
\item Applying machine learning algorithms in order to
|
|
automatically determine the best algorithm as described in
|
|
section \ref{sec:ComMeth}~\parencite{bogason2015ffesl}
|
|
\item Leveraging information gained from prior feature extraction
|
|
to determine aspects such as the signal's noisiness in order to
|
|
select the algorithm best suited to this description.
|
|
\item Dynamic algorithm parameter adoption to improve the likelihood
|
|
of an accurate estimation based on descriptors. Adapting window
|
|
size for example.~\parencite{liuni2012aasas}
|
|
\end{itemize}
|
|
Having determined the optimal set of estimation algorithms and
|
|
selection techniques, these will be implemented in python, or
|
|
potentially a faster compiled language such as C, to create a tool
|
|
capable of creating robust F0 estimations for a range of varying audio
|
|
files.
|
|
|
|
\section{Significance of Research}
|
|
This research aims to explore possible improvements to the overall
|
|
robustness and general accuracy of pitch detection and thus has
|
|
significance in fields such as music and speech analysis and
|
|
transformations. By taking a higher level approach to the problem, it
|
|
is hoped that the careful combination of algorithms will yield a
|
|
superior overall outcome to that of the individual algorithms.
|
|
|
|
\section{Timeline}
|
|
\begin{table}[H]
|
|
\centering
|
|
\label{my-label}
|
|
\begin{tabular}{ll}
|
|
Month 1 - 3 & Initial research into methods and combination techniques \\
|
|
& as well as the set up of initial framework for code if necessary. \\
|
|
Month 3 - 6 & Implementation and testing of individual algorithms. \\
|
|
Month 6 - 9 & Implementation of combination methods. \\
|
|
Month 9 - 12 & Analysis of results and work on method improvements.
|
|
\end{tabular}
|
|
\end{table}
|
|
|
|
\printbibliography
|
|
|
|
\end{document}
|