Software System for Verification and Identification of Speakers

Avatar’s speaker recognition model follows the mechanics of human hearing. Humans do not hear “whole” sounds, they rather perceive certain combinations of “atomic” components for each sound they recognize.

Moreover, the set of “atomic” components for each sound may partially overlap with the set of “atomic” components for other sounds. For example, the set of component structures for sound [a:] overlaps with the set of structures for sound [o:]. Similarly, the sound [i] overlaps with the sound [i:].

With a spectra representation these structures can be interpreted as averaged spectra of the corresponding sounds for each particular speaker. These averaged spectra are then used in the applied model of the Avatar’s Phonemic Machine to distinguish the structural components of individual speech sounds patterns.

The Avatar’s automatic speaker recognition system does not use generally accepted notions such as the pitch frequency and speech formants. Still, these parameters are functions of the averaged integral characteristics of the spectra of “atomic” sounds and are taken into account indirectly in the identification process. They are also present in a number of graphical interfaces of the system.

A significant feature of the Avatar’s voice recognition system is the analysis of short time intervals, from 10 to 30 ms. The actual lengths are determined adaptively, depending on a number of factors. The frequency resolution for calculating spectral characteristics of the “atomic” components is 1 Hz regardless of intervals’ lengths and is achieved using non-orthogonal wavelet transforms based on the Morlet wavelet.

This interpretation of the spectra is fully consistent with the classical concepts, up to the order of resolution in frequency and specific form of time window functions. The Avatar system can analyse voice characteristics of speech recordings that last from 30 to several hundred milliseconds.

At the core of the system is the built-in Phonemic Machine designed on the basis of deep learning neural networks. It analyses each speaker’s individual set of vowel phonemes and compares a wide range of characteristics to other speakers’ speech patterns.

Since speakers are identified by the closeness of the characteristics of their pronunciation of vowels, the analysis is context independent and provides reliable results regardless of speakers’ native language, age, and gender.

The Avatar software system also includes a number of specially built databases that provide solutions for the verification and identification of speakers, depending on the duration of the source phonograms and speaker’s language group.

The basis for the probabilistic identification of speakers is the error graphs of the first and second kind. These error probabilities of the first and second kind are used directly in solving the problems of speaker identification and verification. All probability characteristics were determined using special test sets based on prepared experimental data.

The system has at least an order of magnitude higher frequency resolution when evaluating any parameters of the spectrum of phonogram characteristics compared to existing systems. This is achieved using non-orthogonal spectral transformations.

The effectiveness of the Avatar automatic speech recognition system depends on the length of the source recordings. The feasible audio file duration starts from 30 ms and the longer the audio is, the smaller is the probability of errors.

The probability of errors is calculated based on graphs of errors of the first and second kind for large audio recordings data sets with sampling frequency >= 8000 Hertz.

The use of a special technology for comparing the full spectra of the speakers’ vowel sounds at small time intervals provides highly precise results.

Depending on the audio files lengths the probability of an error for speaker’s identification is 0.2% or less.