![Page 1: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/1.jpg)
Spectral Slope Measurements in Emotionally Expressive Speech
Lucas Tamarit, Martijn Goudbeek& Klaus Scherer
![Page 2: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/2.jpg)
Introduction
• Spectral slope is a measure of voicequality
• Perceptually, voice qualities include harsh, tense, breathy, creaky voice and whisper
• Reflected in the intensity of the harmonicsand more generally in the shape of the power spectrum
![Page 3: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/3.jpg)
Plan
• Overview of existing voice quality indices used in emotion discrimination
• Integration of speaker-dependentinformation in voice quality indices
• Proposition of new indices• Results & discussion• Future plans
![Page 4: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/4.jpg)
Voice quality indices in emotiondiscrimination
• Mostly global, power spectrum orientedmeasures
• Characterizations of the long-term averagespectrum (LTAS)
• Share the idea of splitting the spectrum in a high and a low frequency range (notion of pivot frequency)
![Page 5: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/5.jpg)
Widespread voice quality indices
• Hammarberg Index• Drop-off of the spectral energy above
1000Hz• Relative amount of energy in the high-
(above 1000Hz) versus low-frequencyrange of the spectrum (up to 1000Hz)
![Page 6: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/6.jpg)
Example: The Hammarberg Index
• Defined as the difference between the maximum energy in the 0..2kHz and in the 2..5kHz band
![Page 7: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/7.jpg)
Pivot Frequency
• The lower part of the spectrum mainlycontains phonetic information (F1 and F2)
• The higher part of the spectrum is knownto convey more information about the voice quality
![Page 8: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/8.jpg)
Existing voice quality indices
• Pros– Concise synthesis of the data of whole
speech segments– Directly related to the spectral slope
• Cons– Do not take any speaker dependent
information into account– Rough measures, do not reflect much details
of the spectrum (basically only a slope)
![Page 9: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/9.jpg)
Integrating speaker-dependentinformation
• Replace the fixed 2kHz pivot by a speaker-dependent pivot
![Page 10: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/10.jpg)
Speaker dependent Pivot
• Three approaches– Pitch based– Formants based– Spectral Balance based
• All three approaches are based on neutralreference recordings.
![Page 11: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/11.jpg)
Neutral reference recordings
• For each speaker, we use referencerecordings of the sustained vowel /a/ whenexpressing interest.
• For more information, see Dr. Goudbeek’sposter presentation « Acoustic Profiles in Emotion – The GeMEP Corpus ».
![Page 12: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/12.jpg)
Pitch based pivot
• The fundamental frequency of all reference recordings is extracted thenaveraged.
• Any multiple of that reference frequencycan be used as the pivot.
• Here, we chose the 10th harmonic as the basis for further analysis.
![Page 13: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/13.jpg)
Formants based pivot
• The second formant of all referencerecordings is extracted then averaged.
• The choice of F2 among other formants refers to the aforementionedconsiderations about phonetic and voicequality content.
![Page 14: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/14.jpg)
Spectral balance based pivot
• Defined as the frequency separating the whole spectrum in two parts having a fixedratio of energy λ:
• A value of λ = 50 yields pivot frequenciestypically located between 1kHz and 2kHz.
![Page 15: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/15.jpg)
Spectral Slope Measurements
• Once a pivot has been chosen, severalspectral slope measurements can bederived from it.
• Three approaches :– Modified Hammargberg Index– Spectral Slope– Spectral drop-off curvature
![Page 16: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/16.jpg)
Modified Hammarberg Index
• Same as the original Hammarberg Index, replacing the fixed 2kHz value by the chosen pivot:
• The upper limit is pushed to 8kHz.
![Page 17: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/17.jpg)
Spectral Slope
• Apply simple linear regression to the considered spectral portion
where is the parameter of interest
![Page 18: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/18.jpg)
Spectral drop-off curvature
• Fitting a decaying exponential to the LTAS
(with )
where is the parameter of interest
![Page 19: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/19.jpg)
Summary
• Three pivots– Pitch based– Formants based– Spectral balance based
• Three spectral slope measurements– Modified Hammarberg Index– Spectral Slope– Spectral drop-off curvature
![Page 20: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/20.jpg)
Results
Spectral slope(linear)
Spectral drop-off curvature (decaying
exponential)
HammarbergIndices
• Indices means for all three pivots :
![Page 21: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/21.jpg)
Results
• Significant effect of emotions on all dependent variables (MANOVA on coefficients) : HammI, mHammI, linear, exponential
• Highly similar pattern across pivots• High correlations between linear measures• Exponential least important in discriminant
function
![Page 22: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/22.jpg)
Discussion
• Direct comparison of differentcharacterizations of the LTAS shape of emotionally expressive speech
• Traditional Hammarberg Index shows good performance against new indices propositions
• Spectral slope curvature might not berelated to affect
![Page 23: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/23.jpg)
Future plans
• Compare the present results with analyses on a speaker independent basis.
• Disentangle speaker specific and emotionspecific effects on the shape of the spectrum
![Page 24: Spectral Slope Measurements in Emotionally Expressive Speech](https://reader035.vdocuments.mx/reader035/viewer/2022062216/62a3ef442b573933013c1003/html5/thumbnails/24.jpg)
Thanks for your attention.