The IFA corpus: A phonemically segmented Dutch speech database

R.J.J.H. van Son
Louis C.W. Pols

Affiliation:
The Institute of Phonetic Sciences of the University of Amsterdam
(IFA)
The Amsterdam Center for Language and Communication (ACLC) 
Herengracht 338
NL-1016CG Amsterdam
The Netherlands


The IFA corpus is an open source database of hand-segmented
Dutch speech.  The corpus contains high quality two-channel
speech recordings, using both a head-mounted and a fixed
microphone, from 4 male and 4 female speakers in 8 different
speaking styles:

1.Informal story telling face-to-face to an "interviewer" (I)
2.Retelling a previously read narrative story (R) 

And reading aloud: 

3.A narrative story (T) 
4.A random list of all sentences of the narrative stories (S)
5."Pseudo-sentences" (PS) 
6.Lists of selected words from the texts (W) 
7.Lists of all distinct syllables from the word lists (Sy) 
8.Idiomatic and "diagnostic" sequences (e.g., the 
alphabet, numbers, /hVd/) (Pr)

The total amount of labeled speech is 5:30 hours (50,000 words
and 200,000 segments), speech preparation took around 3
person-weeks per speaker.  Hand segmentation took 1,000 hours
of labeling alltogether. The asymptotic segmentation speed was
about one word, or four boundaries, per minute.  An evaluation
showed that the Median Absolute Difference of the segment
boundaries was 6 ms between labelers, and 4 ms within labelers.
Label differences (substitutions, insertions, and deletions)
were found in 8% of the segments between labelers and 5% within
labelers.

The IFA corpus incorporates some fundamental principles of
Speech Databases.

1) All speech is annotated on several, aligned, linguistic
levels, e.g., from the sentence down to the phoneme level. 
These annotations are strictly hiearchical so they can always
be expressed in a nested XML format.

2) Each and every annotated fragment of speech, e.g., phonemes
or words, has a unique identifier that is used as a pointer to
both the speech fragment (i.e., storage and begin and end
times) and all related data (annotations, acoustic analysis,
perceptive tests).

3) All data is treated in a content neutral way.  Each
annotation is just an aligned ASCII text string (in Praat
TextGrid format) just as each acoustic analysis is treated
identical to the original Speech Waveform.

4) All annotations are amenable to reversible corrections. 
Both "branching" of annotations and retrieving old (original)
annotations are possible.

5) Extensive meta-data on the speakers is collected.

ALL data are stored in a Relational Database and are searchable
with SQL. Each realization of linguistic elements has Parents,
Neighbours, and Children which can be accessed by joining the
relevant tables using the unique fragment pointers. 
Annotations (e.g., POS, Lemma's, ToDI) and measurements (e.g.,
Shadowing delays, reaction times, F0-F3 values) can be added
and removed without affecting access to other data.