Speaker discrimination ability of glottal waveform features

Juan Félix Torres, Elliot MooreInterspeech 2012

TL;DR

This paper investigates how well individual glottal waveform features, extracted from inverse-filtered speech, can distinguish between different speakers. They found that while individual features have moderate success, a combination of these features significantly improves speaker discrimination, even complementing traditional spectral features. This work highlights the underappreciated role of the glottal source in voice identity.

What Problem Does This Paper Address?

Identifying individuals based on their voice is a crucial task in many applications, from security to personalized services. While spectral features (how sound energy is distributed across frequencies) have been widely used for speaker identification, the glottal source – the vibration of vocal folds that initiates speech – also carries unique speaker-specific information. However, extracting reliable glottal features can be challenging due to the complexities and potential artifacts introduced by inverse filtering (a process to isolate the glottal source from the rest of the vocal tract's influence). This paper aims to systematically evaluate the discriminative power of various glottal waveform features, both individually and in combination, to understand their contribution to speaker identification and to see if they can enhance existing spectral-based methods.

Methods

The researchers evaluated 16 different glottal waveform features. These features capture various aspects of the glottal cycle, such as the duration of the open and closed phases of the vocal folds (e.g., Open Quotient - OQ), the speed at which the vocal folds close (e.g., Speed Quotient - SQ), and spectral characteristics of the glottal waveform (e.g., Spectral Tilt - TILT).

The core of their methodology involved a pairwise speaker discrimination task at the frame level. This means they took short segments (frames) of speech and tried to determine if two frames came from the same speaker or different speakers. They used Gaussian Mixture Models (GMMs) for classification, training separate models for each speaker pair. To ensure robustness, they employed a cross-validation strategy, dividing the data into multiple folds and averaging results.

They also compared the performance of glottal features against a set of spectral envelope features (melsub), which are commonly used in speaker identification. Finally, they investigated the benefit of combining glottal and spectral features by creating a combined feature vector.

Datasets

The experiments were conducted using a subset of the TIMIT corpus. Specifically, they selected 50 male and 50 female speakers from the TEST set, using all 10 sentences for each speaker. This provided a substantial amount of data for training and evaluation.

Key Insights and Ideas

  • Glottal features are more than just noise: Despite the challenges in inverse filtering, glottal features demonstrate a significant ability to discriminate between speakers, especially when combined.
  • Gender differences matter: The performance of individual glottal features varied between male and female speakers, suggesting that gender-specific modeling or feature selection might be beneficial.
  • Time-domain features shine: Amplitude-threshold based open and speed quotients (like OQ20, OQ50, SQ20-80) showed particularly strong performance, outperforming LF-model based features and some spectral features for males.
  • Complementary information: Combining glottal features with spectral features led to a statistically significant improvement in speaker discrimination accuracy, indicating that glottal information provides complementary cues to spectral information.

Key Results

  • Individual Male Speaker Discrimination: Most individual glottal features achieved frame-level pairwise speaker discrimination rates between 60% and 68% for male speakers.
  • Individual Female Speaker Discrimination: Female speakers showed slightly lower overall accuracy, with most features scoring above 60% but none exceeding 65%.
  • Combined Glottal Features: A vector of all 16 glottal features achieved significantly higher classification rates, reaching up to 89.4% for males and 86.9% for females.
  • Combined Glottal and Spectral Features: Adding the glottal feature vector to the spectral envelope features (melsub) resulted in a small but statistically significant increase in accuracy (around 0.8% for males and 0.9% for females). This improvement was notable given the already high performance of the spectral features alone (around 95% for males and 93% for females).

Limitations and Future Work

The authors acknowledge that the accuracy of glottal features is influenced by the quality of the inverse filtering process. They also note the inconsistency in feature rankings between genders, suggesting that further research is needed to understand these differences.

Future work could focus on:

  • Identifying a smaller, more discriminative set of glottal features.
  • Developing methods to maximize the complementarity between glottal and spectral features.
  • Investigating how these glottal features perform in more challenging, real-world noisy conditions.

Summary updated

BibTeX
@inproceedings{torres2012speaker,
  author = {Juan Félix Torres and Elliot Moore},
  title = {{Speaker discrimination ability of glottal waveform features}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure