RU2742040C1 - Biometric method for subscriber identification based on speech signal - Google Patents

Abstract

FIELD: physics.

SUBSTANCE: invention relates to computer engineering and communication. Method includes preliminary determination of a set of characteristics of a voice portrait of a subscriber, setting said characteristics for different subscribers, recording them in a database, determination of a speech portrait of the addressed subscriber, comparison of this speech portrait with voice portraits from the database and determination of the subscriber. Determination of the voice characteristics of a subscriber for composing a voice portrait is carried out using simultaneous spectral, time and spectral-time analysis methods together with analogue-to-digital conversion using wavelet transformation. As voice example can be used any fragment of speech, except pauses. Identification is performed not by signal envelope, but by specially processed digital representation thereof. This significantly speeds up the identification process and enables real-time operation on a communication network with a capacity of up to 3000 subscribers.

EFFECT: technical result consists in remote identification of subscriber.

1 cl

Description

The invention relates to the field of information and communication technologies (computers and communications), in particular to an automated system of identification and authentication of a person by biometric and other personality parameters, and is intended to solve the problem of providing access to special information and / or operations with it based on identification and verification by voice, or by voice and face, and through speech recognition.

Known (RU, patent 5995927, publ. 30.11.1999) a method for recognizing a subscriber based on a purely stochastic approach. In this method, subscriber recognition is performed by constructing and comparing covariance matrices of feature descriptions of the input speech signal and speech patterns of known subscribers.

This known method is not suitable for a situation where the used voice messages (passwords) are very short (5 seconds or less), and is also very sensitive to signal loss in certain parts of the speech frequency range due to ambient noise and low-quality microphones.

It is known (RU, patent 2047912, publ. 10.11.1995) a method for recognizing isolated speech words with adaptation to the speaker, based on processing with predistortions of the input speech signal, sampling and sequential segmentation of the speech signal, coding segments with discrete elements, calculating the energy spectrum, measuring formant frequencies and determination of amplitudes and energy in different frequency bands of a speech signal, classification of articulatory events and states, formation and sorting of word standards, calculating distances between word standards with the implementation of a recognized word, making decisions on recognizing or refusing to recognize a word with the addition of a dictionary of standards in the process adaptation to the speaker. The pre-distortion of the input speech signal is performed in the time domain when differentiating with smoothing, the energy spectrum is quantized depending on the dispersion of the communication channel noise, the formant frequencies are determined by finding the global maximum of the logarithmic spectrum and subtracting a given frequency-dependent function from this spectrum, when classifying articulatory events and states determine the fractions of periodic and noise sources of excitation when compared with the threshold of the autocorrelation coefficients of a sequence of rectangular pulses in several frequency bands, the beginning and end of articulatory movements and the corresponding acoustic processes are determined when comparing with the threshold of the likelihood function from the values of the autocorrelation coefficients, formant frequencies and energies in given frequency bands, the speech signal is segmented into intervals between the beginning and end of acoustic processes corresponding to specific articulatory movements, and Consequently, starting with vowel sounds, the segment is identified only if the types of transitions on its left and right boundaries coincide, and segmentation is completed when identifying the pause segments between words on the left and right in time. The standards of words are formed in the form of matrices with binary values of the likelihood of features, and the refusal of recognition is carried out at a normalized difference in the distance from the unknown implementation to the two nearest standards belonging to different words, less than the set threshold.

The disadvantages of this known method for recognizing isolated speech words with adaptation to the speaker is the weak distinctive power of this method when it is used to recognize speakers by pronouncing a voice password, since this method does not distinguish between speakers when they pronounce passwords with the same verbal composition.

Known (RU, patent 2406163, publ. 10.12.2010) is a method for user authentication, containing the stages at which:

- storing a set of personal information in the service program, and this set of personal information obtained from the user during a service-oriented interaction of the user with the service program;

- use the means of extracting information to access the said set of personal information;

- performing an authentication operation that is based on said set of personal information and involves at least one dynamic component, wherein the authentication operation includes sending a question to the user, which is based at least in part on this set of personal information, and the authentication operation is configured in such a way as to invite the user to answer this question in the form of a piece of speech;

- receive a fragment of speech from the user;

- check that the speech fragment is the correct answer to the mentioned question;

- use a computer processor, which is a functional component of the computer, to compare a fragment of speech with a stored voice profile; then providing the user with access to the service program, provided that the speech fragment is the correct answer to the said question, and that the speech fragment at least substantially matches the stored voice profile.

The disadvantage of the known technical solution is the narrow functionality that does not allow the registration and identification of people in various situations, as well as their identification and voice recognition in real time.

The known technical solution was adopted as the closest analogue.

The technical problem, solved by implementing the developed method, consists in providing the possibility of identifying subscribers by voice portrait.

The technical result achieved by implementing the developed method consists in providing the possibility of remote identification of the subscriber. Any fragment of speech except pauses can be used as a voice example.

Identification is carried out not by the envelope of the signal, but by its specially processed digital representation. This significantly speeds up the identification process and makes it possible to work in real time on a communication network with a capacity of up to 3000 subscribers. In the communication channel, information about the speech signal is not transmitted anywhere in the clear - all transmissions in open channels use encryption. This makes it possible to effectively apply this algorithm in cloud services (distributed information and communication systems, where information is stored and processed in different places).

To achieve the indicated technical result, it is proposed to use the developed biometric method for identifying a subscriber by a speech signal. According to the developed method, a preliminary determination of the set of characteristics of the subscriber's speech portrait is carried out, the specified characteristics are established for various subscribers, they are recorded in the database, the speech portrait of the contacting subscriber is determined, this speech portrait is compared with the speech portraits from the database and the subscriber is determined, and the characteristics of the subscriber's voice are determined to compose a speech portrait, a speech portrait is carried out using methods of simultaneous spectral, temporal and spectral-temporal analysis in conjunction with analog-to-digital conversion using Wavelet transform, while the specified characteristics are used, characteristics selected from the group containing at least average value of the spectrum of a speech signal, normalized average values of the spectrum, relative residence time of the signal in the spectral bands, normalized residence time of the signal in the spectral bands, median value of the spectrum speech in bands, the relative power of the speech spectrum in bands, the variation of the envelopes of the speech spectrum, normalized values of the variation of the envelopes of the speech spectrum, the cross-correlation coefficients of the spectral envelopes between the bands of the spectrum, the duration of the phoneme segment, the height of the phoneme segment, the shape factor of the phoneme segment, as well as the characteristics of the tonal sections of the sound waves, non-periodic sections of a sound wave, sections that do not contain speech pauses and the average frequency of pauses. And in the process of pronouncing a speech signal by an unknown subscriber, similar characteristics are determined and by comparing these characteristics with those recorded in the database, the subscriber is determined.

When implementing the developed method, an arbitrary fragment of a speech signal with a duration of 120 sec is converted using a Wavelet transform into a digital code that is stored in a subscriber database. According to the specified fragment, at least the following characteristics of the subscriber's voice signal are evaluated:

The average value of the spectrum of a speech signal is the arithmetic average between the harmonics of the speech signal, representing the extreme upper and extreme lower. It is calculated by processing the code obtained as a result of the Wavelet transform.

Normalized Spectrum Means - Values derived from spectrum averages, normalized to the harmonic mean. It is calculated by processing the code obtained as a result of the Wavelet transform.

The relative residence time of the signal in the spectrum bands is a vector of numerical values corresponding to the time when the signal is within the boundaries of the frequency ranges. The spectrum of the speech signal is localized in the band 100 Hz - 4000 Hz. This strip is divided into uneven components, according to the characteristic color of speech. 8 lanes stand out. In each band, the speech signal is located for a certain time - this is the component of the vector of the relative stay of the signal in the frequency bands. It is calculated by processing the code obtained as a result of the Wavelet transform.

The normalized residence time of the signal in the spectrum bands is a vector value obtained from the vector of the relative residence of the signal in the frequency bands when normalizing the average value of the spectrum of the speech signal. It is calculated by processing the code obtained as a result of the Wavelet transform.

The median value of the speech spectrum in bands is a vector characteristic of the distribution of the amplitude of the speech signal in the bands of the spectrum analysis. It is calculated by processing the code obtained as a result of the Wavelet transform.

The relative power of the speech spectrum in bands is a vector characteristic of the distribution of the speech signal power in the spectrum analysis bands. It is calculated by processing the code obtained as a result of the Wavelet transform.

Variation of the envelopes of the speech spectrum - the vector characteristic is calculated based on the comparison of several speech signals of the same length. It is calculated by processing the code obtained as a result of the Wavelet transform.

Normalized Speech Spectrum Envelope Variation The vector characteristic is calculated on the basis of Speech spectrum envelope variation when normalizing the average spectrum value. It is calculated by processing the code obtained as a result of the Wavelet transform.

Cross-correlation coefficients of spectral envelopes between spectral bands are a vector characteristic for evaluating the coincidence of envelopes. It is calculated by processing the code obtained as a result of the Wavelet transform.

In some embodiments of the developed method, according to a speech fragment 50 sec long, characteristics such as

- Segment duration, phonemes

Segment height

- Segment shape factor

In some embodiments of the developed method, according to a speech fragment in 120 seconds, such voice characteristics are estimated as:

- Coefficient of tonal sections of a sound wave

- Coefficient of non-periodic sections of the sound wave

- Coefficient of sections without speech pauses

average pause frequency

average pause duration

These characteristics are obtained by processing the envelope of the analog speech signal with a specially designed filter.

Each obtained characteristic is assigned a four-digit digital code, resulting in a code sequence of length 64. This sequence is called a speech portrait that uniquely characterizes the subscriber. The speech portrait is stored in the subscriber database. For each new subscriber, a speech portrait is calculated and compared with those in the database. Comparison is made for each four characters separately. If, according to at least 12 characteristics, the value falls within the established tolerance, then a decision is made on the coincidence of speech portraits and the subscriber is identified by the base, and the speech portrait from the recognized signal is added to the base as a clarifying one. In total, for each subscriber, at least three speech portraits are stored, if it is possible to obtain them. Identification is carried out for each of them separately. A positive decision (the subscriber's popularity) is made when two out of three coincide.

An example of implementation.

1. We receive a speech signal.

2. Select arbitrary fragments with a duration of 50 s and 120 s. Fragments can overlap each other.

3. We pass the signals of 50 s and 120 s through a special filter and obtain the values of the characteristics:

Segment duration, phonemes - L1

Phoneme segment height - L2

Phoneme segment shape factor - L3

The coefficient of the tonal sections of the sound wave - L4

Coefficient of non-periodic sections of a sound wave - L5

Coefficient of sections without speech pauses - L6

Average pause frequency - L7

average pause duration - L8

Each value is assigned a four-digit binary code corresponding to the range in which it falls.

A 32 hexadecimal codeword is formed, consisting of the segments K1 - K8 - the acoustic component of the speech portrait (AS RP)

4. A signal segment with a duration of 120 s is passed through the Wavelet transducer and the characteristics are estimated:

Average value of the speech signal spectrum - S1

Normalized Spectrum Means - S2

Relative residence time of the signal in the spectrum bands - S3

Normalized signal dwell time in spectrum bands - S4

Median speech spectrum - S5

The relative power of the speech spectrum in bands - S6

Variation of speech spectrum envelopes - S7

Normalized Speech Spectrum Envelope Variation - S8

Each value of each characteristic fits into the range

0000 - FFFF (hexadecimal). The value is assigned a four-digit hexadecimal code corresponding to the range in which it falls.

A codeword is formed, consisting of 32 hexadecimal symbols of the segments K9 - K16 - the spectral component of the speech portrait (SS RP)

The code words are combined into a 64-bit speech portrait of the RP. For example, we got RP

23AF 4257 AAEE 2433 993F AB01 32F1 FF32 - AS RP

207F 6857 A32E 24A3 453F A351 77F1 F342 - SS RP

The speech portrait is passed through the encoder and transmitted over the communication network.

In the database of speech portraits (DB RP), the received signal is decoded and compared with the RP stored in the database

For example, we took

23AF 4257 AAEE 2433 993F AB01 32F1 FF32 - AS RP

207F 6857 A32E 24A3 453F A351 77F1 F342 - SS RP

And there is a record in the RP DB

23AF 4257 ААЕЕ 993F А 431 32 FA - АС RP

207F 6857A32F 24А3 453F 77FFF F388 - SSRP

This entry has 12 odds that differ by one or overlap.

A decision is made on the identification of the received RP and the received RP is added to the RP DB, as clarifying for the given subscriber.

Claims (1)

A biometric method for identifying a subscriber by a speech signal, including a preliminary determination of a set of characteristics of a speech portrait of a subscriber, establishing the specified characteristics for various subscribers, recording them in a database, determining a speech portrait of a contacting subscriber, comparing this speech portrait with speech portraits from a database and identifying a subscriber, characterized in that the determination of the characteristics of the subscriber's voice to compose a speech portrait is carried out using the methods of simultaneous spectral, temporal and spectral-temporal analysis in conjunction with analog-to-digital conversion using the Wavelet transform, while the characteristics selected from the group are used as these characteristics, containing at least the average value of the spectrum of the speech signal, normalized average values of the spectrum, the relative residence time of the signal in the spectral bands, the normalized residence time of the signal in the specific bands ktra, the median value of the speech spectrum in the bands, the relative power of the speech spectrum in the bands, the variation of the envelopes of the speech spectrum, the normalized values of the variation of the envelopes of the speech spectrum, the cross-correlation coefficients of the spectral envelopes between the spectral bands, the duration of the phoneme segment, the height of the phoneme segment, the shape factor of the phoneme segment, and also the characteristics of the tonal sections of the sound wave, non-periodic sections of the sound wave, sections that do not contain speech pauses and the average frequency of pauses, and in the process of pronouncing a speech signal by an unknown subscriber, similar characteristics are determined and the subscriber is determined by comparing these characteristics with those recorded in the database.