Age recognition based on speech signals using weights supervector

Royi Porat, Dan Lange, Yaniv ZigelInterspeech 2010

TL;DR

This paper introduces a novel method for recognizing a speaker's age from their voice by leveraging Gaussian Mixture Model (GMM) supervectors with Support Vector Machines (SVMs). By focusing on the weights of specific Gaussians within the GMM, the authors achieve improved age classification accuracy compared to existing state-of-the-art methods, particularly for children and seniors.

What Problem Does This Paper Address?

While human listeners can readily estimate a person's age from their voice, automating this process with computers is challenging. Existing research has explored various acoustic features and machine learning models, but accurately distinguishing age groups, especially younger and middle-aged adults, remains difficult. This paper aims to improve age recognition by proposing a new way to represent speaker characteristics derived from GMMs, making them more discriminative for SVM classifiers. The authors also highlight the challenge posed by the lack of large, balanced, and well-defined age-group databases.

Methods

The core of the proposed system lies in a novel approach to creating speaker representations, termed "Gaussian weight supervectors." Instead of the common practice of adapting GMM means, this method focuses on the weights of specific Gaussians within a Universal Background Model (UBM).

Here's a breakdown of the process:

  1. Feature Extraction: Speech signals are pre-processed (e.g., pre-emphasis, down-sampling) and then features like Mel Frequency Cepstral Coefficients (MFCCs) and their derivatives (AMFCCs) are extracted. These features are normalized.
  2. UBM Estimation: A GMM is trained on a large dataset to create a Universal Background Model (UBM). Each Gaussian in the UBM has a unique identifier.
  3. Tokenizer and Weight Supervector Creation: For each speaker:
  4. Likelihood probabilities (scores) are calculated for each speech frame against all Gaussians in the UBM.
  5. For each frame, the top C Gaussians with the highest scores are identified.
  6. A "Gaussian weight" is computed by summarizing the contribution of each Gaussian across all frames for that speaker. This essentially quantifies how much each Gaussian "represents" the speaker.
  7. These weights are concatenated to form a "weight supervector" for the speaker.
  8. SVM Classification: These weight supervectors are then used as input features for an SVM classifier, which is trained to distinguish between different age groups.

The paper contrasts this with the more traditional GMM-UBM adapted means approach, where the average parameters (means) of Gaussians are adapted for each speaker.

Datasets

The proposed system was evaluated on two corpora:

  1. aGender Corpus: A publicly available corpus containing 46 hours of telephone speech. It was divided into training, development, and test sets, with speakers ranging from 7 to 80 years old. For this study, it was divided into four age groups: Child (7-14), Young (15-24), Adult (25-54), and Senior (55-80).
  2. Hebrew Corpus: A new, in-house corpus of Hebrew speakers. It includes 80 male speakers aged 5-96 years. Speakers were recorded pronouncing vowels, words, and free speech. This corpus was divided into four age groups: Children (≤16), Young Adults (17-35), Adults (36-60), and Senior (≥61).

Key Insights and Ideas

  • Discriminative Power of Gaussian Weights: The core hypothesis is that the weights of specific Gaussians within a UBM can be more discriminative for age than the adapted means of those Gaussians. This suggests that how much a particular acoustic pattern (represented by a Gaussian) is present in a speaker's voice is a strong indicator of their age.
  • Simplicity and Efficiency: The proposed weight supervector approach is presented as simpler and potentially more efficient than methods requiring complex dimension reduction techniques often associated with traditional GMM supervectors.
  • Language Dependency: The results suggest that the optimal parameters (number of Gaussians, SVM kernels) can vary depending on the spoken language, indicating that acoustic characteristics related to age might manifest differently across languages.
  • Challenges with Middle Age Groups: Similar to previous work, the system struggled more with classifying young and adult speakers compared to children and seniors. This points to a persistent challenge in age recognition.

Key Results

The proposed weight supervector approach generally outperformed the baseline GMM-UBM adapted means classifier on both corpora.

  • aGender Corpus: Achieved a weighted average recall (WAR) of 56.18% and unweighted average recall (UAR) of 59.45% using a polynomial kernel and 150 best Gaussians. This was better than the baseline's 51.23% UAR.
  • Hebrew Corpus: Achieved a WAR of 56.18% and UAR of 59.45% using a linear kernel and 150 best Gaussians. Again, this surpassed the baseline's performance.
  • Confusion Matrices: Analysis of confusion matrices revealed that children and seniors were classified with high accuracy, while young and adult groups showed poorer performance, consistent across both datasets.

Limitations and Future Work

  • Classification of Young and Adults: The paper acknowledges the ongoing difficulty in accurately classifying young and adult speakers.
  • Future Work: The authors plan to:
  • Investigate additional acoustic features beyond MFCCs to improve classification, particularly for the challenging age groups.
  • Explore other techniques to enhance the overall classification results.
  • Further research the dependency of age recognition on spoken language.

Summary updated

BibTeX
@inproceedings{porat2010age,
  author = {Royi Porat and Dan Lange and Yaniv Zigel},
  title = {{Age recognition based on speech signals using weights supervector}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure