A method of speaker identification based on phoneme mean F-ratio contribution

Songgun Hyon, Hongcui Wang, Chen Zhao, Jianguo Wei, Jianwu DangInterspeech 2012 · OpenReview

Introduction

Speaker identification aims to determine the identity of a speaker regardless of the linguistic content of their speech. Traditional systems predominantly rely on Mel-Frequency Cepstrum Coefficients (MFCCs), which are modeled after the human auditory system's non-linear perception of pitch. However, the Mel-scale is primarily optimized for speech recognition (understanding what is said) rather than speaker recognition (identifying who is speaking). This paper introduces a novel feature extraction method, PMFFCC (Phoneme Mean F-ratio Frequency Cepstrum Coefficients), which adaptively warps frequency filters to prioritize speaker-specific information, achieving a significant improvement in identification accuracy.

The Challenge: Decoupling Identity from Linguistics

A major hurdle in speaker identification is the interference of linguistic information. Standard frequency-warping methods often fail to fully isolate the unique physical characteristics of a speaker's vocal tract from the phonemes being spoken. To address this, the researchers proposed using the "F-ratio"—a measure of the ratio between inter-speaker variance and intra-speaker variance. By calculating this ratio for specific phonemes and then taking a weighted average (the Phoneme Mean F-ratio), the system can effectively "ignore" phonetic changes and focus on individual identity.

Proposed Methodology

The PMFFCC extraction process involves several critical stages designed to enhance speaker-dependent features:

  1. Preprocessing: Includes Voice Activity Detection (VAD) and high-pass filtering to remove noise and silence.
  2. Phoneme-Aware Analysis: The F-ratio is calculated per phoneme across the training set to identify frequency bands that consistently distinguish individuals regardless of the word spoken.
  3. Adaptive Filter Bank Design: Instead of the static triangular filters used in the Mel-scale, the filter bank is warped based on the Phoneme Mean F-ratio. Regions with high speaker-discriminative information receive a higher density of filters (higher resolution).
  4. Feature Generation: The energy spectrum is processed via Discrete Cosine Transform (DCT) to produce 32-order cepstral coefficients.
Figure 1: Block diagram of the proposed PMFFCC feature extraction process.

Dynamic Frequency Resolution

A key innovation is the shift away from the Mel-scale's linear-to-logarithmic resolution. As shown in the comparison of frequency resolutions, the proposed method allocates significantly more resolution to high-frequency bands (above 4 kHz) compared to the standard Mel-scale, which tends to discard detail in these ranges.

Figure 2: Comparison of frequency resolution between the standard Mel-scale and the proposed Phoneme Mean F-ratio method.

Key Insights: The Importance of High Frequencies

The study revealed that speaker-specific information is concentrated in two primary zones:

  • 100 Hz – 300 Hz: Associated with glottal characteristics and fundamental frequency.
  • 4 kHz – 5.5 kHz: Linked to the piriform fossa, a laryngeal cavity that is highly stable and unique to individuals.

Furthermore, the analysis showed that frequencies above 8 kHz—often treated as noise in speech recognition—carry vital speaker information, particularly within consonant sounds. By focusing on these "speaker-rich" zones, the PMFFCC method achieves superior decoupling of identity from speech content.

Experimental Results

The method was tested using the RyongNam2007 Korean speech database, featuring 50 speakers. The results demonstrate that emphasizing frequency bands with high individual dependency significantly reduces error rates.

Feature Set Recognition Rate
MFCC (Standard) 95.06%
FFCC (Standard F-ratio) 97.79%
PMFFCC (Proposed) 98.42%

The PMFFCC method reduced the error rate by 68.02% compared to traditional MFCCs and by 28.5% compared to the standard F-ratio approach.

Conclusion and Future Work

The use of the Phoneme Mean F-ratio to adaptively design filter banks provides a robust framework for speaker identification. While this study was conducted on a Korean corpus, the underlying principle of adaptive frequency warping based on speaker-dependent variance is broadly applicable. Future research will explore the consistency of these discriminative frequency bands across different languages and their performance in noisy, real-world environments.

Summary updated

BibTeX
@inproceedings{hyon2012method,
  author = {Songgun Hyon and Hongcui Wang and Chen Zhao and Jianguo Wei and Jianwu Dang},
  title = {{A method of speaker identification based on phoneme mean F-ratio contribution}},
  booktitle = {Interspeech},
  year = {2012},
  url = {https://openreview.net/forum?id=_E4CW1ITWuPi},
}
Figure