1.0 November l990 (corrections released
January
l991)
Hardware / OS requirements:
Standard SESAM workstation, equipped with Oros AU21 (or AU 22)
DSP board and the library DISK_TRA of the recording software
that accompanies OROS DSP boards
Description:
The software implements an isolated word recogniser.
A word is recorded and endpoint detection is then carried
out by analysing an energy contour and the zero crossing rate.
The endpointed word is preemphasised and windowed by a
Hamming
window.
Discrete Fourier transforms are calculated over the word.
The transform samples are filtered by a filter bank of
triangular filters, linearly distributed in frequency up to 1
kHz and logarithmically distributed above 1 kHz.
From
the filter outputs cepstrum coefficients are
calculated.
The time warping uses eight cepstrum coefficients per
frame, allows local time distortions between 50% and 200% and
is based on a symmetric slope constraint with P=1.
The word chosen is
the word among the reference templates
with the smallest distance to the recorded word.
Developing lab:
Telia Research
Systems Research, Spoken Language Processing
Rudsjoterrassen 2
S-136 80 Haninge
Sweden
Tel: (+46) 8 707 55 72
Fax: (+46) 8 707 55 96
E-mail:gunnar.hult@haninge.trab.se
Contact:
Gunnar Hult, Room A309
WWW Administrator
Fri May 19 11:53:36 MET DST
1995
