Classifications

• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L17/00—Speaker identification or verification
  • G10L17/02—Preprocessing operations, e.g. segment selection; Pattern representation or modelling, e.g. based on linear discriminant analysis [LDA] or principal components; Feature selection or extraction
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L17/00—Speaker identification or verification
  • G10L17/06—Decision making techniques; Pattern matching strategies
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L15/00—Speech recognition
  • G10L15/08—Speech classification or search
  • G10L15/10—Speech classification or search using distance or distortion measures between unknown speech and reference templates
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
  • G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
  • G10L25/27—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 characterised by the analysis technique

Definitions

• the present invention relates to a method and a device for analyzing voice signals.
• the analysis of voice signals notably requires being able to represent a speaker.
• the representation of a speaker by a mixture of Gaussian (“Gaussian Mixture Model” or GMM) is an effective representation of the acoustic or vocal identity of a speaker.
• GMM Gaussian Mixture Model
• this technique it is a question of representing the speaker, in a reference acoustic space of a predetermined dimension, by a weighted sum of a predetermined number of Gaussians.
• This type of representation is precise when we have a large number of data, and there are no physical constraints to store the parameters of the model, nor to execute calculations on these numerous parameters.
• the authors propose to represent a speaker, no longer absolutely in a reference acoustic space, but relatively with respect to a predetermined set of representations of reference speakers also called anchor models, for which there are GMM-UBM models (UBM for "Universal Background Model”).
• GMM-UBM models UBM for "Universal Background Model”
• the proximity between a speaker and the reference speakers is evaluated by means of a Euclidean distance. This greatly reduces the computational loads, but the performances are still limited and insufficient.
• the invention aims to analyze voice signals by representing the speakers with respect to a predetermined set of reference speakers, with a reduced number of parameters reducing the computational loads for time applications real, with acceptable performances, in comparison with an analysis using a representation by the GMM-UBM model.
• the probability density of the similarities between the representation of said speech signals of the speaker ( ⁇ ) and the predetermined set of vocal representations of the reference speakers is represented by a Gaussian distribution ( ⁇ ( ⁇ ⁇ , ⁇ ⁇ )) of vector of mean ( ⁇ ⁇ ) of dimension E and of covariance matrix ( ⁇ ⁇ ) of dimension ExE estimated in the space of resemblances to the predetermined set of E reference speakers.
• information a priori is also introduced into the probability densities of the resemblances ( ⁇ ( ⁇ ⁇ , ⁇ ⁇ j) with respect to the E reference speakers.
• a system for analyzing a speaker's voice signals Comprising databases in which are stored voice of a predetermined set of E reference speakers and their vocal representations associated signals in a predetermined model, as well as audio archive databases, characterized in that it comprises means for analyzing voice signals using a vectorial representation of the similarities between the vocal representation of the speaker and the predetermined set of vocal representations of E reference speakers.
• the databases also store the analysis of the voice signals carried out by said analysis means.
• the invention can be applied to the indexing of audio documents, however other applications can also be envisaged, such as the acoustic identification of a speaker or the verification of the identity of a speaker.
• Other objects, characteristics and advantages of the invention will appear on reading the following description, given by way of nonlimiting example, and made with reference to the single appended drawing illustrating an implementation of a use of the process for indexing audio documents.
• the figure shows an application of the system according to one aspect of the invention for the indexing of audio databases.
• the system comprises means for receiving voice data from a speaker, for example a microphone 1, connected by a connection 2 with or without wire to means 3 for recording a request made by a speaker ⁇ and comprising a set voice signals.
• the recording means 3 are connected by a connection 4 to storage means 5 and, by a connection 6, to acoustic processing means 7 of the request.
• These acoustic processing means transform the voice signals of the speaker ⁇ into a representation in an acoustic space of dimension D by a GMM model of representation of the speaker ⁇ . This representation is defined by a weighted sum of M
• D is the dimension of the acoustic space of the absolute GMM model
• x is an acoustic vector of dimension D, ie vector of cepstral coefficients of a speech signal sequence of the speaker ⁇ in the absolute GMM model
• M denotes the number of Gaussians of the absolute GMM model, generally a power of 2 between 16 and 1024
• the acoustic processing means 7 of the request are connected by a connection 8 to analysis means 9.
• These analysis means 9 are able to represent a speaker by a probability density vector representing the similarities between the vocal representation of said speaker in the chosen GMM model and vocal representations of E reference speakers in the chosen GMM model.
• the analysis means 9 are also able to carry out verification and / or identification tests for a speaker. To carry out these tests, the means of analysis proceed to the elaboration of the vector of probability densities, that is to say similarities between the speaker and the reference speakers. It is a question of describing a relevant representation of a single segment x of the signal of the speaker ⁇ by means of the following equations:
• w ⁇ is a vector of the space of resemblances to the predetermined set of E reference speakers representing the segment x in this representation space
• plx ⁇ ⁇ j is a probability density or probability normalized by a universal model, representing the resemblance of the acoustic representation x ⁇ of a segment of the vocal signal of a speaker ⁇ , knowing a reference speaker ⁇ j ;
• T x is the number of frames or acoustic vectors of the speech segment x;
• ⁇ j j is a probability representing the resemblance of the acoustic representation x ⁇ of a voice signal segment of a speaker ⁇ , knowing a reference speaker ⁇ j ;
• p ( ⁇ ⁇ u BM ) is a probability representing the resemblance of the acoustic representation x ⁇ of a voice signal segment of a speaker ⁇ in the UBM world model;
• M is the number of Gaussians of the relative GMM model, generally power of 2 between 16 and 1024;
• D is the dimension of the acoustic space of the absolute GMM model;
• x ⁇ is an acoustic vector of dimension D, ie vector of cepstral coefficients of a speech signal sequence of the speaker ⁇ in the absolute GMM model;
• b k (x) represents, for k ⁇ l to D, Gaussian dens
• ⁇ ⁇ represents components of the vector of average ⁇ ⁇ of dimension E of the resemblances ⁇ ( ⁇ ⁇ , ⁇ ⁇ J of the speaker ⁇ with respect to E reference speakers
• ⁇ ". represents components of the covariance matrix ⁇ ⁇ of dimension ExE of the sets ⁇ ( ⁇ ⁇ , ⁇ ⁇ j of the speaker ⁇ with respect to the E reference speakers.
• the analysis means 9 are connected by a connection 10 to learning means 11 making it possible to calculate the vocal representations , in the form of vectors of dimension D, E reference speakers in the GMM model chosen.
• the learning means 11 are connected by a connection 12 to a database 13 comprising voice signals from a predetermined set of speakers and their associated voice representations in the GMM reference model.
• the database 13 is connected by the connection 14 to the analysis means 9 and by a connection 15 to the acoustic treatment means 7.
• the system further comprises a database 16 connected by a connection 17 to the acoustic treatment means 7 , and by a connection 18 to the analysis means 9.
• the database 16 includes audio archives in the form of vocal articles, as well as the associated vocal representations in the GMM model chosen.
• the database 16 is also able to store the associated representations of the audio articles calculated by the analysis means 9.
• the learning means 11 are further connected by a connection 19 to the acoustic processing means 7.
• the learning module 11 will determine the representations in the GMM reference model of the E reference speakers by means of the voice signals of these E reference speakers stored in the database 13, and of the acoustic processing means 7. This determination takes place according to relations (1) to (3) mentioned above.
• This set of E reference speakers will represent the new acoustic representation space.
• These representations of the E reference speakers in the GMM model are stored in memory, by example in database 13. All of this can be done offline.
• the acoustic processing means 7 calculate a vocal representation of the speaker in the predetermined GMM model as explained previously in reference to relations (1) to (3) above.
• the acoustic processing means 7 have calculated, for example offline, the vocal representations of a set of S test speakers and a set of T speakers in the predetermined GMM model. These sets are separate. These representations are stored in the database. 13.
• the analysis means 9 calculate, for example offline, a vocal representation of the S speakers and T speakers compared to the E reference speakers.
• This representation is a vector representation with respect to these E reference speakers, as described above.
• the analysis means 9 also perform, for example offline, a vocal representation of the S speakers and T speakers compared to the E reference speakers, and a vocal representation of the articles of the speakers from the audio database.
• This representation is a vector representation with respect to these E reference speakers.
• the processing means 7 transmit the voice representation of the speaker ⁇ in the predetermined GMM model to the analysis means 9, which calculate a voice representation of the speaker ⁇ .
• This representation is a probability density representation of the resemblances to the E reference speakers. It is calculated by introducing information a priori to the voice representations of T speakers. Indeed, the use of this a priori information makes it possible to keep a reliable estimate, even when the number of available speech segments of the speaker ⁇ is small.
• We introduce information a priori by means of the following equations:
• ⁇ ⁇ vector of mean of dimension E of resemblances ⁇ ( ⁇ ⁇ , ⁇ ⁇ J of speaker ⁇ with respect to E reference speakers;
• N ⁇ number of segments of voice signals from speaker ⁇ represented by N ⁇ vectors of the space of resemblances to the predetermined set of E reference speakers;
• ⁇ ⁇ vector of mean of dimension E of the resemblances ⁇ ( ⁇ ⁇ , ⁇ ⁇ ) of the speaker ⁇ with respect to the E reference speakers, with introduction of information a priori;
• ⁇ ⁇ covariance matrix of dimension ExE of the resemblances ⁇ ( ⁇ ⁇ , ⁇ ⁇ ) of the speaker ⁇ with respect to
• the analysis means 9 will compare the vocal representations of the request and of the articles of the base articles of the base by tests in identification and / or verification of the speakers.
• the speaker identification test consists in evaluating a likelihood measure between the vector of the test segment w x and the set of representations of the articles in the audio base.
• the speaker verification test consists in calculating a likelihood score between the vector of the test segment w x and the set of representations of the articles of the audio base normalized by its likelihood score with the representation of the information a priori.
• the segment is authenticated if the score exceeds a given predetermined threshold, said score being given by the following relation:
• This invention can also be applied to other uses, such as recognition or identification of a speaker.
• This compact representation of a speaker makes it possible to drastically reduce the cost of computation, because there are much less elementary operations in view of the drastic reduction in the number of parameters necessary for the representation of a speaker. For example, for a request for 4 seconds of words from a speaker, that is to say 250 frames, for a GMM model of dimension 27, to 16 Gaussian the number of elementary operations is reduced by a factor of 540 , which greatly reduces the computation time.
• the memory size used to store the representations of the speakers is significantly reduced. The invention therefore makes it possible to analyze the vocal signals of a speaker by drastically reducing the computation time and the storage memory size of the vocal representations of the speakers.

Applications Claiming Priority (1)

Publications (1)

Family

ID=34130575

Family Applications (1)

Country Status (7)

Families Citing this family (9)

Family Cites Families (17)

• 2003
  • 2003-07-01 CN CNA038267411A patent/CN1802695A/zh active Pending
  • 2003-07-01 US US10/563,065 patent/US7539617B2/en not_active Expired - Fee Related
  • 2003-07-01 AU AU2003267504A patent/AU2003267504A1/en not_active Abandoned
  • 2003-07-01 WO PCT/FR2003/002037 patent/WO2005015547A1/fr active Application Filing
  • 2003-07-01 KR KR1020067000063A patent/KR101011713B1/ko not_active IP Right Cessation
  • 2003-07-01 JP JP2005507539A patent/JP4652232B2/ja not_active Expired - Fee Related
  • 2003-07-01 EP EP03748194A patent/EP1639579A1/de not_active Withdrawn

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events