How TANDEM-STRAIGHT works (visualization)

*** Under construction ***

by Hideki Kawahara (Wakayama University, Japan)

This page shows visualization movies to illustrates principles of operations working behind TANDEM-STRAIGHT system.

Spectral envelope extraction

Periodic pulse train

Level axis of the power spectrum is linear.

Synthetic vowel /a/

Level axis of the power spectrum is logarithmic.

Real speech example

Japanese vowel sequence /aiueo/, spoken by a male speaker.

(***Fixed spectrogram position***) Japanese vowel sequence /aiueo/, spoken by a male speaker.

Please use slider of the movie controller for inspecting how TANDEM and STRAIGHT works.

Source information extraction by XSX (eXcitation Structure eXtractor)

Real speech example

How dedicated Periodicity detector subsystem of XSX works.

Integration of individual dedicated detectors

Number of channels per octave=1, (sharpness of loglag weight) bw=1

Number of channels per octave=3, (sharpness of loglag weight) bw=2.8

Number of channels per octave=3, (sharpness of loglag weight) bw=1

Periodicity salience landscape

Periodicity salience landscape with F0 candidates. Top 5 candidates are color coded based on salience ranking.

Response to test signals

Periodic pulse train (137 Hz) plus white Gaussian noise with different S/Ns.,
0 dB: from 0 ms to 100 ms
10 dB: from 100 ms to 200 ms
20 dB: from 200 ms to 300 ms
30 dB: from 300 ms to 400 ms
40 dB: from 400 ms to 500 ms
50 dB: from 500 ms to 600 ms
noise only: from 600 ms to 700 ms

Periodic pulse train (137 Hz) pluse white Gaussian noise with different S/Ns.,
Following optional parameter setting was used:
opt.framePeriod = 1;
opt.setLagPeriodicityMapForF0 = 1;
opt.numberOfHarmonicsForExtraction = 8;
opt.f0ceil = 1000;
opt.channelsPerOctave = 6;

Systematic test signals

Modulation of alternating pulse amplitude

Modulation of alternating pulse location

Direct link to components

Last update: Sat Sep 18 15:14:52 JST