<chapter>
<title>The Tilt Intonation Model</title>
<toc depth=3></toc>
<para>
<emphasis>Tilt</emphasis> is a phonetic model of intonation that
represents intonation as a sequence of continuously parameterised
events.
The tilt library is a set of functions which analyses, synthesizes and
manipulates tilt representations.
</para>
</sect1>
<sect1><title id="tilt-overiew">Theoretical Overview</title>
<para>
The basic unit in the tilt model is the <emphasis>intonational
event</emphasis>. Events occur as instants with nothing between them,
as opposed to segmental based phenomena where units occur in a
contiguous sequence. The basic types of intonational event are
<emphasis>pitch accents</emphasis> and (following the popular
terminology) <emphasis>boundary tones</emphasis>. Pitch accents
(denoted by the letter a) are F0 excursions associated with
syllables which are used by the speaker to give some degree of
emphasis to a particular word or syllable. In the tilt model, boundary
tones (b) are rising F0 excursions which occur at the edges of
intonational phrases and as well as giving the hearer a cue as to the
end of the phrase, can also signal effects such as continuation and
questioning. A combination event ab occurs when a pitch accent
and boundary tone occur so close to one another that only a single
pitch movement is observed. There are different kinds of pitch accents
and boundary tones: the choice of pitch accent and boundary tone
allows the speaker to produce different global intonational tunes
which can indicate questions, statements, moods etc to the hearer.
</para><para>
<FIGURE ID="tiltfig03" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="../arch_doc/tilt_fig03.gif" FORMAT="GIF"></GRAPHIC>
</FIGURE>
<xref linkend="tiltfig03"> shows a Schematic representation of F0,
intonational event relation and segment relation in the Tilt
model. The linguistically relevant parts of the F0 contour, which
correspond to intonational events, are circled. The events, labelled a
for pitch accent and b for boundary are linked to the syllable nuclei
of the syllable relation. Note that every event is linked to a
syllable, but some syllables do not have events. </para><para>
Unlike traditional intonational phonology schemes \cite{ph:thesis},
\cite{tobi} which impose a categorical classification on events, Tilt
uses a set of continuous parameters. These parameters, collectively
known as <emphasis>tilt parameters</emphasis>, are determined from
examination of the local shape of the event's F0 contour.
</para><para>
The tilt model is built on a simpler model, the rise/fall/connection (RFC) model.
</para><para>
In the RFC model, each event is modelled by a rise part followed by a
fall part. Each part has an amplitude and duration, and two parameters
are used to give the time position of the event in the utterance and
the F0 height of the event. <xref linkend="tiltfig01"> shows a typical
pitch accent with these parameters marked.
</para><para>
<FIGURE ID="tiltfig01" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="../arch_doc/tilt_fig01.gif" FORMAT="GIF"></GRAPHIC>
</FIGURE>
The RFC parameters for an utterance are therefore:
<itemizedlist>
<listitem><para>rise amplitude (Hz)</para></listitem>
<listitem><para>rise duration (seconds)</para></listitem>
<listitem><para>fall amplitude (Hz)</para></listitem>
<listitem><para>fall duration (seconds)</para></listitem>
<listitem><para>position (seconds)</para></listitem>
<listitem><para>F0 height (Hz)</para></listitem>
</itemizedlist>
</para><para>
Sometimes events don't have rise or fall parts, and in these cases the
amplitude and duration of the missing part is set to 0. The position
parameter can be specified in two ways: either as the distance from
the start of the utterance, or the distance from the start of the
vowel of the associated syllable. The latter is more linguistically
meangingful, but as vowel boundaries are not always available, the
former is often used.</para><para>
</para><para>
While the RFC model can accurately describe F0 contours, the mechanism
is not ideal in that the RFC parameters for each contour are not as
easy to interpret and manipulate as one might like. For instance there
are two amplitude parameters for each event, when it would make sense
to have only one. </para><para>
</para><para>
The <emphasis>Tilt</emphasis> representation helps solve these
problems by transforming the four amplitude and duration RFC
parameters into three Tilt parameters:
<itemizedlist>
<listitem><para>amplitude (Hz): the sum of the magnitudes of the rise and fall amplitudes. </para></listitem>
<listitem><para>duration (seconds): the sum of the rise and fall durations. </para></listitem>
<listitem><para>tilt: a dimensionless number which expresses the overall
<emphasis>shape</emphasis> of the event, independent of its amplitude or
duration. </para></listitem>
</itemizedlist>
The position and F0 height parameters are the same as before.
</para><para>
The tilt representation is superior to the RFC representation in that
it has fewer parameters without significant loss of
accuracy. Importantly, it can be argued that the tilt parameters are
more linguistically meaningful.
</para>
<para>
In describing the tilt model, we use the term
<emphasis>analysis</emphasis> to describe the process of producing a
tilt representation from an F0 contour, and <emphasis>synthesis
</emphasis> to describe the process of prodcing a F0 contour from a
tilt representation.
</para>
</sect2>
<sect2><title>RFC Analysis</title>
<sect3><title>Locating Events in the F0 contour</title>
<para>
The first stage in analysis is to find the intonational events in an
F0 contour. EST does not directly provide a means for doing this. In
practice this is either done by hand by a human labeller, or
automatically by the HMM auto event labeller. The current HMM event
labeller is based on the HTK system and hence can't be part of EST,
but an outline of the system follows:
</para><para>
The automatic event detector uses continuous density hidden Markov
models to perform a segmentation of the input utterance. A number of
units are defined and a HMM is trained on examples of that kind from a
pre-labelled training corpus using the Baum-Welch algorithm
\cite{baum:72}. Each utterance in the corpus is acoustically processed
so that it can be represented by sequence of evenly spaced
frames. Each frame is a multi-component vector representing the
acoustic information for the time interval centred around the frame.
</para><para>
Recognition is performed by forming a network comprising the HMMs for
each unit in conjunction with an n-gram language model which gives the
prior probability of a sequence of n units occurring. To perform
recognition on an utterance, the network is searched using the
standard Viterbi algorithm to find the most likely path through the
network given the input sequence of acoustic vectors.
</para><para>
It is our intention to put a complete event labeller in EST in the future.
</para>
</sect3><sect3 id="ov-rfc-analysis"><title>Producing an RFC representation from an utterance's
events and F0 contour</title>
<para>
An utterance's events are represented in a relation. Initially, events
are stored as regions with start and stop times as this is the most
common output format of labellers (both human and automatic).
</para><para>
For example, for utterance kdt_016, a set of basic event labels is as
follows (in xlabel format):
<screen>
0.290 146 sil
0.480 146 c
0.620 146 a
0.760 146 c
0.960 146 a
1.480 146 c
1.680 146 a
1.790 146 sil
</screen>
</para><para>
Events are labelled "a", and silences "sil". The use of the "c" label
is to allow start times which differ from the end of the previous
event. Conceptually, this can alsow be represented as follows:
<screen>
name:sil start:0.0 end:0.290
name:a start:0.290 end:0.620
name:a start:0.760 end:0.960
name:a start:1.480 end:1.680
name:sil start:1.790 end:1.790
</screen>
</para><para>
The other component for analysis is the utterance's F0 contour, which
is stored in a track. The contour must be continuous (i.e. have no
breaks), and its frames must be specified at fixed intervals. For best
performance the contour should have been smoothed.
</para><para>
The RFC analysis component takes the approximate labels and the
smoothed F0 contour, fits rise and fall shapes, and hence determines
an optimal set of RFC parameters for the utterance.
</para><para>
For each event, a peak picking algorithm decides if the event has a
rise part only, a fall part only or a rise part followed by a fall
part.
</para><para>
<FIGURE ID="tiltfig02" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_fig02.gif" FORMAT="GIF"></GRAPHIC>
</FIGURE>
For each part, a search region, shown in <xref linkend="tiltfig02">,
is defined around the approximate start and end boundaries as defined
in the input label file. The search region is controlled by a number
of parameters:
</para>
<itemizedlist>
<listitem><para>start_limit: the distance in seconds before each input start
boundary that the start search region should begin.</para></listitem>
<listitem><para>end_limit: the distance in seconds after each input end
boundary that the end search region should begin.</para></listitem>
<listitem><para>range: the end and beginnings of the start and end regions
respectively, specified as a fraction of the overall label duration.
</para></listitem>
</itemizedlist>
<para>
For example, a pitch accent starts at 1.45 seconds and ends at 1.75
seconds. If the start and end limit are both defined to be 0.1 seconds
and the range is 0.4 (40%), then the start region starts at 1.35
seconds and ends at 1.55, and the end region starts at 1.65 and ends
at 1.85. The matching algroithm will synthesize every possible shape
lying within this region, measure the distance between each and the
actual contour, and pick the one with the lowest distance.
</para><para>
The final results of the matching process is a relation of events,
each with the 6 RFC parameters are descibed above.
</para>
<para>
The program <ulink linkend="progtiltanalysis"><command>tilt_analysis</command></ulink> will perform RFC matching
given a label file and F0 contour. The function
<function>rfc_analysis</function> takes a F0 contour, a relation and a
set of options and returns the RFC parameters in the features of each
item in the relation.
</para>
</sect3>
</sect2><sect2 id="rfc2tilt"><title>RFC to Tilt Conversion</title>
<para>
The rise and fall RFC parameters can be converted to Tilt parameters
using the following equations.
<emphasis>Amplitude</emphasis> is the sum of te magnitudes of the rise
and fall amplitudes:
</para>
<EQUATION ID="tilt_eq01" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq01.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
<para>
<emphasis>Duration</emphasis> is the sum of the of the rise and fall durations:
</para>
<EQUATION ID="tilt_eq02" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq02.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
<para>
<emphasis>Tilt</emphasis> can be measured with respect to amplitude:
</para>
<EQUATION ID="tilt_eq03" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq03.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
<para>
or duration:
</para>
<EQUATION ID="tilt_eq04" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq04.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
<para>
The tilt model assumes that these are strongly correlated so that an
average of the two is representative of the shape of the event:
</para>
<EQUATION ID="tilt_eq05" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq05.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
<para>
</para><para>
The is no stand alone program to do this conversion, but the
<command>tilt_analysis</command> can do this conversion in addition to
performing the RFC matching as described above.
</para><para>
The function <function>rfc_to_tilt</function> takes a relation
containing RFC parameterised items and converts it to a relation
containing Tilt paramterised items.
</para><para>
Another function, also called <function>rfc_to_tilt</function> takes a
Features object containing the 4 rise fall paramaters and writes the 3
tilt paramaters into another features object. This function can be
used to do rfc_to_tilt conversion for a single event.
</para>
</sect2><sect2 id="tilt2rfc"><title>Tilt to RFC Conversion</title>
<para>
The Tilt parameters can be converted to RFC parameters using the
follwing equations, which are rearranged from <xref
linkend="tilt_eq01"> to <xref linkend="tilt_eq05">:
<EQUATION ID="tilt_eq06" FLOAT="0">
<TITLE>Rise amplitude
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq06.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
<EQUATION ID="tilt_eq07" FLOAT="0">
<TITLE>Fall amplitude
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq07.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
<EQUATION ID="tilt_eq08" FLOAT="0">
<TITLE>Rise duration
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq08.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
<EQUATION ID="tilt_eq09" FLOAT="0">
<TITLE>Fall duration
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq09.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
</para><para>
The is no stand alone program to do this conversion, but the
<command>tilt_synthesis</command> can do this conversion in addition to
generating a F0 contour.
</para><para>
The function <function>tilt_to_rfc</function> takes a relation
containing Tilt parameterised items and converts it to a relation
containing RFC paramterised items.
</para><para>
Another function, also called <function>tilt_to_rfc</function> takes a
Features object containing the 3 Tilt paramaters and writes the 4 rise
fall RFC paramaters into another features object. This function can be
used to do tilt_to_rfc conversion for a single event.
</para>
</para>
</sect2><sect2 id="ov-rfc-to-tilt"><title>RFC to F0 Synthesis</title>
<para>
An F0 contour can be generated from a set of RFC parameters using the
follwing equations.
Events are generated as piecewise combinations of quadratic functions:
<EQUATION ID="tilt_eq10" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq10.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
</para><para>
Between events, straight lines are used:
<EQUATION ID="tilt_eq11" FLOAT="0">
<TITLE>
</TITLE>
<GRAPHIC FILEREF="arch_doc/tilt_eq11.gif" FORMAT="gif"></GRAPHIC>
</EQUATION>
</para><para> The stand alone program
<command>tilt_synthesis</command> can do this conversion. It takes a
RFC label file as input and produces a F0 file. This program can also
generate a F0 file directly from a Tilt label file</para><para>
The function <function>rfc_synthesis</function> takes a relation
containing RFC parameterised items and produces a F0 contour in a
Track. </para><para>
The function <function>synthesize_rf_event</function> takes a Features
object containing the 4 rise fall RFC paramaters and generates the F0
contour for a single event.
</para>
</para>
</sect2>
</sect1>
<sect1><title>Exectuable Programs</title>
<simplesect><title><link
linkend="tilt-analysis-manual">tilt_analysis</link></title><para>
Produces a Tilt or RFC analysis of a F0 contour, given a set label
file containing a set of approximateintonational event boundaries.
</para></simplesect>
<simplesect><title><link
linkend="tilt-synthesis-manual">tilt_synthesis</link></title><para>
tilt_synthesis generates a F0 contour, given a label file containing
parameteriszed Tilt or RFC events. </para></simplesect>
Generation of F0 contours can be done with:
<simplesect><title><link
linkend="pda-manual">pda</link></title><para>
</para></simplesect>
</sect1>
&docpp-Tiltfunctions;
</chapter>
<!--
Local Variables:
mode: sgml
sgml-doctype: "speechtools.sgml"
sgml-parent-document: ("speechtools.sgml" "chapter" "book")
End:
-->
