TL;DR
This paper investigates how a speaker's age influences their acoustic space, a key component in speaker identification systems. By explicitly modeling age as a factor within established factor analysis models (Eigenvoice and Inter-speaker Variability), the authors demonstrate that incorporating age information can lead to improved speaker identification performance, even with a small number of age-related factors.
What Problem Does This Paper Address?
Automatic speaker identification systems aim to recognize individuals based on their unique voice characteristics. These systems often rely on models that capture the underlying "acoustic space" of a speaker, representing their voice with a reduced set of latent factors. While models like Eigenvoice and Inter-speaker Variability have proven effective, they primarily focus on general speaker variations and channel distortions. However, observable speaker traits like gender, dialect, and crucially, age, also play a significant role in shaping a person's voice. Existing systems often overlook or implicitly handle age, leading to potential performance degradation, especially when training data is not age-balanced. This paper addresses the gap of explicitly understanding and leveraging the impact of age on the speaker acoustic space for more robust speaker identification.
Methods
The core of this work lies in adapting existing factor analysis models to explicitly incorporate age.
- Eigenvoice Model Enhancement: The standard Eigenvoice model represents a speaker's acoustic space with a matrix
vand their specific coordinates with a vectory. This paper proposes splitting thevmatrix into two components:va(age-dependent) andvn(age-independent). The model then becomess = va * ya + vn * yn, whereyarepresents the speaker's coordinates in the age acoustic space. This allows for a direct modeling of age-related voice characteristics. - Inter-speaker Variability Model Consideration: For the Inter-speaker Variability model, which includes common factors (
vy) and specific factors (dz), age cannot be directly incorporated into the matrixddue to its fixed dimension. Instead, the authors propose designing the data pool used to traindwith a specific age distribution to indirectly account for age. - Training and Estimation: Maximum Likelihood re-estimation is used to train the model parameters. For the age-dependent Eigenvoice model, a modified Baum-Welch statistic is employed to estimate
vaandvnseparately, ensuring independence between the age-dependent and age-independent components.
Datasets
- Training Data:
- Switchboard II Phase 2 and 3
- Switchboard Cellular Part 1 and 2
- Fisher English Corpus
- NIST 2004 and 2005 SRE enrollment data
- Evaluation Data:
- NIST 2008 Speaker Recognition Evaluation (SRE) corpus (telephone-telephone condition, male trials only).
The paper highlights that many common corpora (like Switchboard and Fisher) have an unbalanced age distribution, with a majority of speakers in the 20-40 year range, leading to a shortage of data for older speakers.
Key Insights and Ideas
- Age as a Distinct Acoustic Factor: The research confirms that age is not just a demographic attribute but a discernible factor that influences the acoustic properties of a speaker's voice, contributing to their unique acoustic space.
- Explicit Age Modeling Improves Performance: By explicitly modeling age within the factor analysis framework, even with a small number of age-specific factors, a tangible improvement in speaker identification performance can be achieved. This suggests that current models might be leaving valuable information untapped.
- Data Pool Balancing Matters: For models where direct age factor incorporation is difficult (like the Inter-speaker Variability model), carefully curating the training data pool to achieve a balanced age distribution can significantly impact performance, especially for underrepresented age groups.
- Age-Dependent vs. Age-Independent Components: The separation of the acoustic space into age-dependent and age-independent components provides a more nuanced understanding of speaker variability.
Key Results
- Eigenvoice Model: When incorporating 2 age factors into the Eigenvoice model (with a total of 300 factors) and applying channel compensation, the Equal Error Rate (EER) was reduced from 10.06% to 9.90%. This represents a relative improvement of 1.16% EER, achieved by dedicating only a small fraction (2/300) of the speaker acoustic space to age.
- Inter-speaker Variability Model: Evaluating different age-balanced data pools for training the diagonal matrix
dshowed that a data pool specifically focused on older speakers (though smaller in size) yielded the best performance (8.81% EER) compared to using all speakers (9.07% EER) or a younger speaker-focused pool (9.13% EER). This underscores the importance of age distribution in the training data.
Limitations and Future Work
- Age Segmentation: The study uses a simplified age segmentation into two classes (younger/older) for practical reasons, acknowledging that more granular age information could potentially yield further improvements.
- Gender Specificity: The evaluation was conducted on male trials only due to the unbalanced age distribution in the training data. Future work could explore age's role in female speakers.
- Complexity vs. Performance: While the improvements are noted, the paper acknowledges that the relative gain from adding age factors might seem small. However, it emphasizes the importance of this aspect for researchers aiming to build more robust systems.
- Future Directions: The authors suggest that further research into developing age-sensitive Eigenvoice models and exploring more sophisticated age segmentation techniques could lead to even greater performance enhancements in speaker identification systems.
Summary updated