TL;DR
This paper introduces an age-dependent approach to estimate children's height and normalize their speech using subglottal resonances (SGRs). By recognizing that children's vocal tract properties change significantly around age 11, the authors developed age-specific models that improve both height estimation accuracy and speech recognition performance compared to existing methods.
What Problem Does This Paper Address?
Estimating a speaker's physical characteristics, like height, from their voice has applications in forensics and speaker identification. Similarly, normalizing speech for automatic speech recognition (ASR) systems is crucial for handling variations in speakers' voices. While these techniques have been explored for adult speech, children's speech presents a unique challenge due to rapid developmental changes in their vocal tracts. Specifically, the acoustic properties of children's voices change considerably as they grow, particularly around puberty (around age 11). Previous work on speaker normalization and height estimation using subglottal resonances (SGRs) – vibrations in the airway below the vocal folds – has primarily focused on adults. This paper aims to bridge this gap by developing age-dependent models for children's speech, recognizing that a one-size-fits-all approach is insufficient.
Methods
The core of this paper lies in leveraging subglottal resonances (SGRs), specifically the first three (Sg1, Sg2, Sg3), for both height estimation and speaker normalization.
- Age-Dependent SGR Estimation: The authors first analyze the relationship between SGRs and formants in children's speech. They observe that children can be broadly categorized into two age groups: below 11 years and above 11 years, with distinct acoustic properties. Based on this, they train separate regression models to estimate Sg1, Sg2, and Sg3 for each age group. These models are derived using a cross-validation approach on their dataset.
- Height Estimation: Building on the established inverse correlation between SGRs and speaker height in adults, the paper proposes an age-dependent height estimation algorithm for children. They use linear regression models to predict height from the estimated SGRs.
- Speaker Normalization: For speaker normalization, the paper adapts a linear frequency warping technique. They specifically focus on using the age-dependent Sg3, as initial experiments suggest it yields the best results. The normalization process involves warping the test speaker's SGRs to match a reference speaker's SGRs, with adjustments made using a maximum likelihood approach.
Datasets
The primary dataset used in this study is the WashU-UCLA child corpus. This corpus contains simultaneous microphone and subglottal accelerometer recordings from 46 child speakers aged 6-17 years. The speakers are divided into two age groups: 6-11 years (24 speakers) and 11-17 years (22 speakers). For speaker normalization experiments, the TIDIGITS database was used. This database consists of adult speech for training ASR models and children's speech for testing.
Key Insights and Ideas
- The 11-Year-Old Cutoff is Crucial: The paper strongly supports the idea that children's vocal tract acoustics change significantly around age 11, making age-dependent modeling essential for accurate height estimation and speaker normalization.
- Sg3's Importance: While Sg1 and Sg2 have been explored for children's speech before, this work highlights the significant role of Sg3, particularly for speaker normalization, due to its strong correlation with vocal tract length.
- SGRs as Robust Features: Subglottal resonances prove to be robust features for children's speech, showing strong correlations with both physical height and vocal tract characteristics relevant for ASR normalization.
- Age-Dependent Models Outperform: The experiments clearly demonstrate that using age-specific SGR estimation models leads to better performance in both height estimation and speaker normalization compared to age-independent models.
Key Results
- Height Estimation Accuracy: The age-dependent height estimation algorithm achieved mean absolute height estimation errors of less than 3.8cm for younger children (under 11) and 4.9cm for older children (over 11).
- SGR Estimation Accuracy: Average RMS errors for estimating Sg1, Sg2, and Sg3 using the age-dependent model were 51 Hz, 128 Hz, and 168 Hz, respectively.
- Speaker Normalization Performance: Using age-dependent Sg3 for linear frequency warping resulted in statistically significant word error rate reductions of up to 26% compared to conventional Vocal Tract Length Normalization (VTLN) on the TIDIGITS dataset. The best performance was achieved using oracle age information, but F3-based age estimation also yielded competitive results.
Limitations and Future Work
- Limitations:
- The paper acknowledges that while F3 can be used to estimate age groups for normalization, it's not a perfect measure and can only roughly separate speakers.
- The height estimation algorithm relies on the inverse correlation between SGRs and height, which might have limitations in extreme cases or with specific vocal pathologies.
- Future Work:
- The authors plan to evaluate the effectiveness of their algorithm on a larger database.
- They also intend to further investigate age estimation for children's speech using SGRs.
Summary updated