Age verification using a hybrid speech processing approach

Ron M. Hecht, Omer Hezroni, Amit Manna, Ruth Aloni-Lavi, Gil Dobry, Amir Alfandary, Yaniv ZigelInterspeech 2009

TL;DR

This paper tackles the challenge of automatically verifying a speaker's age group using speech. They propose a novel hybrid approach that combines traditional acoustic features with higher-level lexical (word-based) information, achieving a significant reduction in verification error compared to either method alone. This work highlights that both the sound of speech and the words used carry distinct, complementary age-related information.

What Problem Does This Paper Address?

The ability to automatically determine a person's age from their speech is an interesting problem with potential applications in areas like personalized services, content filtering, and even forensic analysis. While previous research has largely focused on analyzing the acoustic properties of speech (how it sounds), this paper argues that the content of speech – specifically, a person's vocabulary – also provides valuable clues about their age. The challenge lies in effectively combining these different types of information to build a robust age verification system.

Methods

The authors propose a hybrid approach that integrates two distinct systems:

  • Acoustic Age Estimation System: This system uses a Gaussian Mixture Model (GMM) and Support Vector Machine (SVM) framework. It extracts acoustic features (Mel-Frequency Cepstral Coefficients or MFCCs) and uses GMMs to model speaker characteristics. The core idea is to adapt these GMMs and then use SVMs to classify age groups based on these acoustic supervectors. This is a lower-level analysis, focusing on the physical production of speech.
  • Word-based Age Estimation System: This system operates at a higher level, analyzing the linguistic content. It first transcribes speech into words using a large-vocabulary continuous speech recognition (LVCSR) system. Then, it extracts "N-gram" features, which represent the probability of word sequences (unigrams for individual words, bigrams for pairs of words, etc.). SVMs are then used to classify age groups based on these word-level features.

The hybrid approach combines the outputs of these two systems. They experimented with two merging strategies: simply summing the log-likelihood ratio (LLR) scores from each system, and using a linear SVM combiner. The intuition behind the hybrid approach is that the errors made by the acoustic and word-based systems are likely uncorrelated, meaning that if one system makes a mistake, the other is less likely to, leading to a more accurate overall decision.

Datasets

  • LDC Fisher English Part I and II Corpora: These large, diverse corpora were used for both training and testing. The Fisher corpora are noted for their extensive collection of telephone conversations from thousands of distinct speakers.
  • Approximately 11,000 conversations from Fisher Part II were used for training.
  • Over 8,000 conversations from Fisher Part I were used for testing.
  • Conversations from speakers present in both training and testing sets were excluded from testing to ensure independence.

Key Insights and Ideas

  • Vocabulary as an Age Indicator: The paper strongly demonstrates that a person's choice of words and word sequences (their vocabulary) is a significant indicator of their age. This moves beyond purely acoustic analysis.
  • Orthogonality of Acoustic and Lexical Information: The authors hypothesize and show evidence that the acoustic and word-based age indicators are largely uncorrelated. This means they capture different aspects of age-related speech changes, making their combination particularly powerful.
  • Hybrid Approach Superiority: Combining acoustic and lexical information significantly outperforms either approach individually, suggesting that a holistic view of speech (both sound and content) is crucial for accurate age verification.
  • Age-Related Speech Changes at Multiple Levels: The paper reinforces the idea that age affects speech production at various stages, from the initial message formulation (vocabulary choice) to the physical vocal tract and vocal cord mechanisms (acoustic properties).

Key Results

  • Acoustic System Performance: Achieved an average Equal Error Rate (EER) of 28.0%.
  • Word-based System Performance: Achieved an average EER of 28.7% using the lexical-based approach.
  • Hybrid System Performance: The combined approach significantly reduced the verification error to an average EER of 24.1%.
  • N-gram Feature Performance: Word bigrams and pair unigrams generally performed better than simple unigrams, especially as the vocabulary size (M) increased.

Limitations and Future Work

  • Limitations:
    • The paper focuses on a four-age-group verification task. The performance for finer-grained age estimation or binary age classification (e.g., child vs. adult) is not explicitly detailed.
  • The "hybrid SVM word unigram" results are shown in Figure 3, but the paper mentions that the difference between the two merging techniques (summing scores vs. linear SVM combiner) was small, implying that the specific merging method might not be a critical factor.
  • Future Work:
    • The authors suggest that the uncorrelated nature of the two systems is key to their success, implying that further exploration of combining different types of uncorrelated information could lead to even better performance.
  • While not explicitly stated as future work, the success of the hybrid approach naturally leads to considering more sophisticated fusion techniques or exploring other higher-level linguistic features.

Summary updated

BibTeX
@inproceedings{hecht2009age,
  author = {Ron M. Hecht and Omer Hezroni and Amit Manna and Ruth Aloni-Lavi and Gil Dobry and Amir Alfandary and Yaniv Zigel},
  title = {{Age verification using a hybrid speech processing approach}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure