On using voice source measures in automatic gender classification of children's speech

Gang Chen, Xue Feng, Yen-Liang Shue, Abeer Alwan2010 · DOI

TL;DR

This paper investigates the effectiveness of voice source acoustic features, specifically Cepstral Peak Prominence (CPP), Harmonic-to-Noise Ratio (HNR), and a spectral tilt measure (H₂-H₄), for automatically classifying the gender of children's speech. The findings demonstrate that incorporating these voice source features significantly improves classification accuracy compared to using only traditional vocal tract features like fundamental frequency (F0) and formant frequencies, particularly for pre-adolescent children.

What Problem Does This Paper Address?

Automatically distinguishing between male and female children's voices is more challenging than for adults. This is because the physiological differences in the glottis and vocal tract that create distinct acoustic signatures in adults are less pronounced in children and are still developing. While previous research has explored vocal tract features (like F0 and formants) and some voice source features for adult gender classification, a systematic investigation into the utility of specific voice source measures for children's speech was lacking. This paper aims to fill that gap by evaluating the contribution of CPP, HNR, and H₂-H₄ to gender classification accuracy in children across different age groups.

Methods

The researchers employed a Support Vector Machine (SVM) classifier with a Radial Basis Function kernel for gender classification. They defined several feature sets:

  • M0 (Baseline): Included vocal tract features (F0, F1, F2, F3, B1, B2).
  • M1: M0 plus two voice source measures from prior work (H₁-H₂ and H₁-A*).
  • M2: M0 plus Cepstral Peak Prominence (CPP).
  • M3: M0 plus CPP and Harmonic-to-Noise Ratio (HNR).
  • M4: M0 plus CPP, HNR, and H₂-H₄ (a measure of mid-frequency spectral tilt).

These feature sets were used to train and test the SVM classifier. For each experiment, 70% of the data was used for training and the remaining 30% for testing, with utterances from a particular speaker used exclusively for either training or testing. The performance was also compared against traditional Mel-Frequency Cepstral Coefficients (MFCCs) using a Gaussian Mixture Model (GMM) classifier.

Datasets

The study utilized the CID database, which contains speech data from 3418 utterances spoken by 174 male and 140 female speakers across five age groups: 8-9, 10-11, 12-13, 14-15, and 16-17 years old. The data consisted of recordings of specific vowel sounds.

Key Insights and Ideas

  • Voice Source Matters for Kids: Unlike adults where vocal tract features are often sufficient, voice source characteristics become crucial for differentiating children's genders, especially before puberty.
  • CPP as a Key Indicator: Cepstral Peak Prominence (CPP), a measure related to breathiness, showed a noticeable increase in its differentiating power between male and female speakers in older age groups (10-11 and 12-13).
  • HNR's Surprising Trend: The study found that HNR was generally higher for females than males across age groups, which is counterintuitive given that females are often perceived as breathier (and breathiness is associated with lower HNR). This suggests a complex interaction between acoustic properties and perceptual voice quality that warrants further investigation.
  • Mid-Frequency Tilt Adds Value: The H₂-H₄ measure, reflecting mid-frequency spectral tilt, also contributed to improved classification accuracy, indicating it captures gender-specific spectral characteristics.
  • Puberty as a Turning Point: The most significant accuracy improvements were observed around the 12-13 age group, coinciding with the onset of puberty in males and females, suggesting that developmental changes in voice production are particularly impactful at this stage.

Key Results

  • The addition of CPP, HNR, and H₂-H₄ (feature set M4) consistently improved gender classification accuracy across all age groups compared to the baseline (M0) and the previous best set (M1).
  • Feature set M4 achieved an average improvement of 4.4% over the baseline (M0) and 3% over M1.
  • A substantial improvement of 8% was observed for the 12-13 age group when comparing M4 to M0.
  • For the 8-9 age group, the classification accuracies remained relatively low (<61%), with minimal improvement from adding voice source features, suggesting these features are less discriminative in younger children.
  • Interestingly, classification accuracy was consistently higher for females than for males across all age groups, regardless of the feature set used.

Limitations and Future Work

  • Limited Age Range: The study focused on children and pre-adolescents. The effectiveness of these features for older adolescents and adults was not directly assessed in this paper.
  • Vowel-Specific Analysis: The analysis was primarily conducted on steady-state portions of specific vowels. The authors suggest that future work should focus on applying these voice source measures to continuous speech.
  • HNR Anomaly: The unexpected trend of higher HNR in females requires further exploration to understand the underlying acoustic and perceptual reasons.
  • Perceptual Correlates: While CPP is linked to breathiness, the paper hints at complex interactions between acoustic measures and perceptual voice quality that could be further investigated.

Summary updated

Abstract

Acoustic characteristics of speech signals differ with gender due to physiological differences of the glottis and the vocal tract. Previous research [1] showed that adding the voice-source related measures H ∗ 1 − H ∗ 2 and H ∗ 1 − A ∗ improved gender classification accuracy compared to using only the fundamental frequency (F0) and formant frequencies. H ∗ i refers to the i–th source spectral harmonic magnitude, and A ∗ refers to the magnitude of the source spectrum at the i–th formant. In this paper, three other voice source related measures: CPP, HNR and H ∗ 2 − H ∗ 4 are used in gender classification of children’s voices. CPP refers to the Cepstral Peak Prominence [2], HNR refers to the harmonic-to-noise ratio [3], and H ∗ 2 − H ∗ 4 refers to the difference between the 2nd and the 4th source spectral harmonic magnitudes. Results show that using these three features improves gender classification accuracy compared with [1]. Index Terms: gender classification, gender identification, voice source
BibTeX
@misc{chen2010using,
  author = {Gang Chen and Xue Feng and Yen-Liang Shue and Abeer Alwan},
  title = {{On using voice source measures in automatic gender classification of children's speech}},
  year = {2010},
  doi = {10.21437/interspeech.2010-251},
}
Figure