Speaker Recognition: Feature Extraction

1. Short-Term Spectral Features

常用的有MFCC, LPCC, LSF, PLP。实际应用中，如何选择哪个特征参数，重要性不如如何做好channel compensation。

 

2. Voice Source Features

常用的有Fundamental frequency, glottal pulse shape。

与“Short-Term Spectral Features”的比较如下：

“Based on the literature, voice source features are not as discriminative as vocal tract features but fusing these two complementary features can improve accuracy”

Experiments also suggest that the amount of training and testing data for the voice source features can be significantly less compared to the amount of data needed for the vocal tract features (10 seconds vs 40 seconds in [188]). A possible explanation
for this is that vocal tract features depend on the phonetic content and thus require sufficient phonetic coverage for both the training and test utterances. Voice source features, in turn, depend much less on phonetic factors.

 

3. Spectro-Temporal Features

A common way to incorporate some temporal information to features is through 1st and 2nd order time derivative estimates, known as delta and doubledelta  coefficients.

 

4. Prosodic Features

"Prosody refers to non-segmental aspects of speech, including for instance syllable stress, intonation patterns,
speaking rate and rhythm. One important aspect of prosody is that, unlike the traditional short-term spectral features, it spans over long segments like syllables, words, and utterances and reflects differences in speaking style, language background, sentence type, and emotions to mention a few."  TBD

 

5. High-Level Features

"Speakers differ not only in their voice timbre and accent pronounciation, but also in their lexicon - the kind
of words the speakers tend to use in their conversations."  TBD

 

 参考：

[1] An Overview of Text-Independent Speaker Recognition: from Features to Supervectors