Voice source characterisation Gerrit Bloothooft UiL-OTS Utrecht University.

Voice source characterisation

Gerrit Bloothooft

UiL-OTS Utrecht University

Emasters School Leuven 2002 Voice Source Characterization 2

Voice research

To describe and model the properties of the vocal sound source from view points of:– Physiology– Acoustics– Perception


Importance of the voice

• Speech synthesis– Towards natural sounding synthesis

• Speech recognition– Using source properties in recognition

• Speaker recognition/identification– Voice source characteristics are essential

• Diagnosis– Pathologies, voice classifications


Voice possibilities

Limited use of voice in speech• Range of the fundamental

frequency• Vocal intensity range• Spectral variation


Focus in this presentation

How do acoustic voice source characteristics vary as a functionof F0 and vocal intensity


Voice profile measurement

Thirties: Intensity range as function of various pitches– manual measurement

Eighties: Automatic computation ofF0 and Intensity– computer measurement– visual feedback– additional parameters


Measurement unit

• One decibel• One semi-tone


Measurement procedure

• Subject in front of computer screen• Microphone on head set (30 cm)• Just phonate, sing, and see the result

immediately

• Best results with recording protocol• Feed back stimulates extreme

phonations


Fundamental frequency (Hz)

Voca

l In

ten

sity

(d

B S

PL)

Sam

ple

den

sity

Voice profile / density



Voca

l In

ten

sity

(d

B S

PL)

Sam

ple

den

sity

Voice profile / speech area


Acoustic voice quality parameters

• Jitter– Stability of periodicity– Asymmetry in vocal folds

• Crest factor– Max amplitude divided by average

energy– Relates to spectral slope

• Many more …


Crest factorV

oc a

l In

ten

sity

(d

B S

PL)


Cre

st f

act

or


Jitter


Vo

cal

inte

nsi

ty (

dB

SP

L)

regular

irregular


Real time presentation

Screen presentation• One data point per F0-I cell

Advanced data storage [new]• Full audio signal • Full distribution of data per F0-I cell

• Data for screen presentation


Advantages

• Reusability of recordings

• Statistical analysis per F0-I cell

• Study of time-varying behavior


Crest factorV

oc a

l In

ten

sity

(d

B S

PL)


Cre

st f

act

or


Median smoothing of crest factor

Voc a

l In

ten

sity

(d

B S

PL)


Cre

st f

act

or

Crest factor median smoothed


Vocal Registers

Different movement patterns of the vocal folds

• Pulse register (creaky voice)• Modal register• Falsetto register


Pulse register

• Less than 50 Hz• Irregular • Long closed period


Fundamental Frequency (Hz)

Voc

al I

nten

sity

(dB

SP

L)

Pulse register


Modal register

• “Normal” use of voice• Active role of M. Vocalis• Vocal folds thick and completely

vibrating

• Wide range in F0 and intensity

• Flat spectrum



Voc

al I

nten

sity

(dB

SP

L)

Modal register


Falsetto register

• Higher pitches• M. Vocalis passive, tense vocal

ligaments through M.Cricothyroidus

• Edge vibration of vocal volds• Sound poor in higher harmonics (in

untrained subjects)



Voc

al I

nten

sity

(dB

SP

L)

Falsetto register



Voc

al I

nens

ity

(dB

SP

L)

Register overlap


Chest- en head voice

Refer to secundary vibratory sensations in the body

• Chest voice: loud modal register• Head voice:

– males: higher, softer modal register in overlap area with falsetto register

– women: falsetto register



Voc

al I

nten

sity

(dB

SP

L)

Chest voice and Head voice

chest

head


Registers and voice profiles

With a description using

• Iso-crest factor lines• Iso-jitter lines


Iso-crest factor lines

4 dB

6 dB

Vo c

al I

nten

sity

(dB

SP

L)

Cre

st f

acto

r



Vo c

al I

nten

sity

(dB

SP

L)


3 %

Jitt

er (

%)

Iso-jitter lines


New representation

• Areas defined by iso-parameter lines– crest factor < 4 dB– crest factor > 4 dB, < 6 dB– crest factor > 6 dB– jitter < 3 %– [relative rise time < 6 %]


Areas in the phonetogramV

o cal

Int

ensi

ty (

dB S

PL

)


Jitter > 3%, unstable

RRT < 6 %pressed-like Crest factor < 4 dB

sine-like

Emasters School Leuven 2002 Voice Source Characterization 33Fundamental frequency (Hz)

Vocal registers in the phonetogram

Falsettoupper boundary

Modallower boundary

Chest voiceboundary

Vo c

al I

nten

sity

(dB

SP

L)


Comparison of voice profiles

Characterisation of

• Voice pathologies• Voice classifications

Reuse stored voice profiles of subjects with known voice history


Important features

• Contour has limited value– but most research goes into that

direction (norm profiles)

• Distribution of acoustical parameters across the voice profile tells much more


• Unit for comparison

Voice profile unit defined by small range of F0 and Vocal Intensity

• Distributions of acoustic voice parameters per unit

Probability density function per parameter• Model

Hidden Markov Model

We need


IN OUT

two unconnected states per phonetogram unit

• vocal registers• start and end of phonetion

Unit model


Speech Voice Profile

• phoneme model F0/I unit model

• not labeled labeled by F0 and I

• spectral envelope acoustic voice parameters• language model unrestricted transitions

“forced alignment

recognition”

Correspondences


Crest factor distributions

training subject 1

0

500

4 5 6 7 8 9 10 11 12 13 14 15

test subject 1

0

500

4 5 6 7 8 9 10 11 12 13 14 15

training subject 2

0

500

4 5 6 7 8 9 10 11 12 13 14 15

test subject 2

0

500

4 5 6 7 8 9 10 11 12 13 14 15



Voc

al I

nten

sity

(dB

SP

L)

Dis

tinc

tive

ness

Most distinctive states


Conclusions

• Voice profiles can enhance our understanding of vocal behaviour in a visually attractive way

• Current data storage opens a series of important research topics

• Market opportunities for “light” versions

Voice source characterisation Gerrit Bloothooft UiL-OTS Utrecht University.

Documents

emasters school leuven

voice classifications

vocal intensity slide

source properties

voice speech synthesis

vocal sound source

screen presentation

spectral variation slide