Age estimation from telephone speech using i-vectors

Mohamad Hasan Bahari, Mitchell McLaren, Hugo Van hamme, David van LeeuwenInterspeech 2012

TL;DR

This paper introduces a novel approach to estimating a speaker's age from telephone speech by leveraging i-vectors, a compact representation of speech utterances. By replacing traditional GMM supervectors with i-vectors and using Support Vector Regression (SVR), the authors achieve improved age estimation accuracy compared to existing methods, demonstrating the power of i-vectors for this task.

What Problem Does This Paper Address?

Estimating a person's age from their voice, known as speaker age estimation, has numerous practical applications, from personalizing customer service to aiding in forensic investigations. However, it's a challenging task due to several factors: perceived age often differs from chronological age, obtaining large, age-balanced speech datasets is difficult, and speech itself contains significant variability unrelated to age (like emotional state or physical characteristics). Previous methods often relied on high-dimensional GMM mean supervectors, which are computationally expensive and can struggle with limited data. This paper aims to improve the accuracy and efficiency of age estimation by exploring a more compact and robust speech representation.

Methods

The core innovation of this paper lies in the application of i-vectors for age estimation. I-vectors are a low-dimensional representation derived from GMM supervectors using a technique called total variability modeling. This modeling assumes that a GMM mean supervector ($\mu$) can be decomposed into a universal background model (UBM) mean ($m$) and an utterance-dependent offset ($m + Tw$), where $T$ represents a low-dimensional subspace and $w$ is the i-vector.

The proposed system follows these steps:

  1. I-vector Extraction: For each speech utterance, an i-vector is extracted.
  2. Support Vector Regression (SVR): The extracted i-vectors are then used as input features for an SVR model, which is trained to predict the speaker's age.

The paper also investigates two session compensation techniques, Linear Discriminant Analysis (LDA) and Within-Class Covariance Normalization (WCCN), to mitigate variations in speech due to factors like microphone type, channel, or gender. These techniques aim to reduce within-speaker variability and enhance between-speaker differences.

Datasets

The authors utilize speech data from the National Institute for Standard in Technology (NIST) 2010 and 2008 Speaker Recognition Evaluations (SRE) databases. These databases are chosen for their large number of speakers and the availability of age-related metadata. Specifically, the core protocols of these databases are used for training and testing. The development dataset used to train the total variability subspace and UBM includes over 30,000 speech recordings from NIST 2004-2006 SRE, LDC Switchboard 2, and Switchboard Cellular.

Key Insights and Ideas

  • I-vectors are a powerful alternative to GMM supervectors: The paper strongly suggests that i-vectors, due to their compact nature and ability to capture essential speaker characteristics, offer a more sophisticated and effective representation for age estimation than traditional GMM mean supervectors.
  • Dimensionality Reduction Matters: The success of GMM-PCA-SVR and GMM-WPPCA-SVR over GMM-SVR highlights the importance of reducing the dimensionality of GMM supervectors. I-vectors inherently provide this dimensionality reduction.
  • Session Compensation is Tricky for Age: While LDA and WCCN are effective for speaker recognition, they did not yield improvements for age estimation in this study. This suggests that age, being a continuous and ordinal variable, might require different compensation strategies than categorical tasks like speaker identification. The authors hypothesize that the large intra-class variation and the ordinal nature of age are key challenges.
  • Acoustic Feature Choice Impacts Performance: Using more detailed acoustic features (MFCC60D) generally led to better performance than simpler ones (MFCC26D) for the baseline systems.

Key Results

  • The proposed i-vector-SVR system significantly outperforms baseline methods like GMM-SVR, GMM-PCA-SVR, and GMM-WPPCA-SVR.
  • The i-vector-SVR system achieved a Mean Absolute Error (MAE) of 7.61 years for female speakers and 7.63 years for male speakers on the NIST datasets. This represents an improvement over the best baseline systems.
  • The paper notes that the i-vector-SVR system achieved an estimation accuracy improvement of 5.94%, 2.93%, and 2% relative to GMM-SVR, GMM-WPPCA-SVR, and GMM-PCA-SVR, respectively, simply by using i-vectors instead of GMM mean supervectors.
  • Session compensation techniques (LDA and WCCN) applied to i-vectors did not improve the MAE, with i-vector-LDA-SVR and i-vector-WCCN-SVR showing slightly worse performance than the direct i-vector-SVR.

Limitations and Future Work

  • Session Compensation Challenges: The authors acknowledge that LDA and WCCN were not beneficial for age estimation. They suggest that further research is needed to develop session compensation methods that account for age being an ordinal number and the significant intra-class variation present in speech.
  • Data Requirements for Compensation: The limited data might have hindered the effectiveness of LDA and WCCN, as discriminant subspace directions may not emerge clearly from estimation noise.
  • Future Work: The paper implicitly suggests exploring alternative session compensation techniques tailored for continuous, ordinal properties like age. Further investigation into how to best model the ordinal nature of age within these compensation frameworks would be a valuable next step.

Summary updated

BibTeX
@inproceedings{bahari2012age,
  author = {Mohamad Hasan Bahari and Mitchell McLaren and Hugo Van hamme and David van Leeuwen},
  title = {{Age estimation from telephone speech using i-vectors}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure