Takagi KAZUYUKI (高木 一幸)

In most of current speech processing techniques, MFCC obtained from amplitude spectrum and delta-MFCC calculated as time derivative of MFCC are widely used as acoustic features. However, these features consider neither frequency derivative of amplitude spectrum nor phase information of speech waveform. Local feature and group delay spectrum are among the features claimed by previous works to possess such information useful for speech processing. We therefore examine their effectiveness on speech recognition performance. We conducted phoneme recognition experiments using speaker-dependent phoneme HMMs trained with local feature, group delay spectrum, and MFCC in same speaker, same gender, and different gender conditions. We obtained highest recognition rate by local feature, while the other features showed better performance for some phonemes. Likelihood combination of local feature, group delay spectrum, and MFCC HMMs yielded better phoneme recognition rate than the case in which each HMM was used solely. Results show that it is promising that recognition performance degradation can be alleviated by a combination of local feature, group delay spectrum, and MFCC.

In most of current speech processing techniques, MFCC obtained from amplitude spectrum and delta-MFCC calculated as time derivative of MFCC are widely used as acoustic features. However, these features consider neither frequency derivative of amplitude spectrum nor phase information of speech waveform. Local feature and group delay spectrum are among the features claimed by previous works to possess such information useful for speech processing. We therefore examine their effectiveness on speech recognition performance. We conducted phoneme recognition experiments using speaker-dependent phoneme HMMs trained with local feature, group delay spectrum, and MFCC in same speaker, same gender, and different gender conditions. We obtained highest recognition rate by local feature, while the other features showed better performance for some phonemes. Likelihood combination of local feature, group delay spectrum, and MFCC HMMs yielded better phoneme recognition rate than the case in which each HMM was used solely. Results show that it is promising that recognition performance degradation can be alleviated by a combination of local feature, group delay spectrum, and MFCC.

Wesetarticulatoryclassesbasedonarticulatoryfeature,andweuseposteriorprobabilitiesonarticula-toryclassesforlanguageidenti cation.PosteriorprobabilityoneacharticulatoryclassiscalculatedbyarticulatoryfeatureextractorbasedonGMMs.Theposteriorprobabilityvaluesofthearticulatoryclassesareconcatenatedtoformvectorateachanalysisframe.ThesevectorsarethenquantizedtoyieldVQcodesequence,whichisusedasthetrainingdataforan-gramlanguagemodel.Theresultsoflanguageidenti cationexperimentbetweenJapaneseandEnglishshowedchangeinidenti cationperformancebycodebooksize.Themethodthatuseslanguage-dependentarticulatoryfeatureextractorshowedidenti cationrateof98.1%whencodebooksizewas64,andthemethodthatuseslanguage-independentarticulatoryfeatureextractorshowedidenti cationrateof95.6%whencodebooksizewas256.

Extraction of features from input speech that are effective in distinguishing the language is a key issue for language identification system. We use posterior probabilities on articulatory classes as features for language identification. Posterior probability on each articulatory class is calculated by GMMs. Each GMM is trained with MFCC data of speech segments labeled with the phonemes or acoustic events that correspond to the articulatory class. The posterior probability values of the articulatory classes are concatenated to form an articulatory-feature-class-posterior-probability (AFCPP) vector at each analysis frame. These vectors are then quantized to yield VQ code sequence, which is used as the training data for a n-gram language model. Language identification is performed by selecting the n-gram model that yields the highest likelihood for the AFCPP vector sequence of the input utterance. Language identification experiment between Japanese and English by the present method showed identification rate of 97.1%.