UCL Young Researchers 2005

This page provides supplementary information
for the paper:

Ian S Howard & Mark A Huckvale, "Training a Vocal Tract Synthesizer to Imitate Speech using Distal Supervised Learning", Submitted to Specom 2005.

A presentation based on the work in this paper is available in PowerPoint format here.

Speech data produced by the babble generator

Babble generator run with different phonetically motivated sampling of the vocal tract space.
The durations between the states are drawn from a Gaussian distribution. with the following parameters:
fast: mean =150ms, sd = 50ms
slow: mean = 300ms , sd = 100ms

1a. Fast babble vowel space	2a. Fast /baba/ babble	3. Fast babble Vowel and Consonant Space
vowelSpaceFast	AbASpaceFast	AiubAiuSpaceFast
1b. Slow babble vowel space	2b. Slow /baba/ babble	3. Slow babble Vowel and Consonant Space
vowelSpaceSlow	AbASpaceSlow	AiubAiuSpaceSlow

Re-synthesized speech spoken by a male subject

Utterance: ' Boogie boogie ba ba ba ba, boogie boogie ba ba ba' spoken by a male subject

Real Speech Input	Via Direct Inverse Model	Via Distal retrained Inverse Model
First Attempt: realSpeechIAH3	First Attempt: resynthSpeechIAH3
Improved: boogierealSpeech	Improved: boogiedirInv	Improved: boogiedistInvBab

Re-synthesized speech sung by a male subject

Utterance: The song Daisy sung by a male subject:
Daisy Daisy,
give me your answer do.
I'm half crazy,
all for the love of you.
We won't have a stylish marriage,
I can't afford a carriage,
but you'll look sweet,
upon the seat,
of a bicycle made for two.

Real Speech Input	Via Direct Inverse Model	Via Distal retrained Inverse Model
daisySung2realSpeech	daisySung2dirInv	daisySung2distInvBab