ï~~Proceedings of the International Computer Music Conference (ICMC 2009), Montreal, Canada
August 16-21, 2009
SYNTHESIZING GHOST VOICES IN GHOST WINDS TALKING
Lydia Ayers
Tower 14, Flat 7A
HKUST
Clear Water Bay, Kowloon
HONG KONG
ABSTRACT
This research describes an additive synthesis design with
timbre morphing for the ghost voices in Ayers' Ghost
Winds Talking. The design captures many of the subtle
timbral and expressive characteristics of the spoken voices
of four famous composers: Harry Partch, Lou Harrison,
Nadia Boulanger and Aaron Copland. The flexible design
models the dynamic spectra of the voices in Csound, and
although the speech sounds like unintelligible English, it is
very suitable for the ghost voices used in the composition.
1. INTRODUCTION
One day the wind blew across the open bathroom window
as though it were playing harmonics on a giant flute (see
Figure 1). Ayers set a microphone in the bathroom and
recorded the sounds. Ghost Winds Talking combines the
edited wind sounds with ghost voices synthesized from
recordings of four famous composers: Harry Partch, Nadia
Boulanger, Aaron Copland and Lou Harrison. We don't
know whether anyone else can recognize their vocal
timbres, but perhaps that is appropriate for ghosts.
Andrew Homrner
Department of Computer
Science and Engineering
HKUST
Clear Water Bay, Kowloon
HONG KONG
harmonics, Ayers chose some samples which she
nicknamed the "staccato whistle" and the "Chinese opera
soprano." The "Chinese opera soprano" was especially
expressive filtered and fluttered. Extra sampled noises
used for percussive effect included amplified rain clicks
and a car door slam, which turned out to sound rather
similar. The singing, whistling and general moaning of the
wind reminded her of ghost voices, and inspired her to
contrast those sounds with the ghosts of real composers'
voices.
1.2. The Ghost Voices
We analyzed speech samples from four famous composers:
Harry Partch, Nadia Boulanger, Lou Harrison and Aaron
Copland, to synthesize for the ghosts. Although speech is
pitched, we didn't separate the timbres according to their
frequencies, as we have in previous instrument modelling
projects [1, 3, 4, 5, 6, 10]. Instead, we based our voice
design on our previous didgeridoo synthesis design [7],
and, as the didgeridoo doesn't typically play scales, we
had focused that research on didgeridoo sounds on a single
pitch. We then chose a collection of interesting sounds to
analyze for the instrument model.
2. USABLE FEATURES OF THE PREVIOUS
ADDITIVE SYNTHESIS DESIGN
We found that the additive synthesis model used in our
previous work was a good starting point for this project,
but it needed some refinement to produce the timbre
morphing of the voices. The following summarizes the
features that we reused in the new model.
The design uses 63 sine waves to model the rich spectra
of the voice timbres (see Figure 2). Each code block
produces a sine wave at the required harmonic frequency,
with a slight random inharmonicity. Each harmonic has a
large variation in its spectral average amplitude which
makes the tones sound more lifelike and unique. The
attack and decay times become longer as the harmonics get
higher, which makes the overall tone get brighter as it gets
louder and less bright as it decays. Each harmonic's
amplitude envelope is multiplied by a slow random noise
and a fast random noise (jitter) and the harmonics are
Figure 1. The bathroom window.
1.1. The Wind Samples
The basic samples used in the piece were the wind samples
blowing across the bathroom window. In addition to the
basic whistling, which produced amazing glissandi and
331
