Texas Instruments Corporation

Yu-Hung Kao

Yifan Gong

Kazuhiro Kondo

Barbara J Wheatley

Ikuo Kudo

A small vocabulary speech recognizer suitable for implementation on a 16-bit fixed-point DSP is described. The input speech xt is sampled at analog-to-digital (A/D) converterand the digital samples are applied to MFCC (Mel-scaled cepstrum coefficients) front end processing. For robustness to background noises, PMC (parallel model combination)is integrated. The MFCC and Gaussian mean vectors are applied to PMC. The MFCC and PMC provide speech features extracted in noise and this is used to modify the HMMs. The noise adapted HMMs excluding mean vectors are applied to the search procedure to recognize the grammar. A method of computing MFCC comprises the steps of: performing dynamic Q-point computation for the preemphasis, Hamming Window, FFT, complex FFT to power spectrum and Mel scale power spectrum into filter bank steps, a log filter bank step and after the log filter bank step performing fixed Q-point computation. A polynomial fit is used to compute log2 in the log filter bank step. The method of computing PMC comprises the steps of: computing noise MFCC profile, computing cosine transform MFCC into mel-scale filter bank, converting log filter bank into linear filter bank with an exponential wherein to compute exp2 a polynomial fit is used, performing a model combination in the linear filter bank domain; and converting the noise compensated linear filter bank into MFCC by log and inverse cosine transform.

Implementing a high accuracy continuous speech recognizer on a fixed-point processor

A typical English pronunciation dictionary takes up to 1,826,302 bytes in ASCII to store. A five times compression while maintaining computability is achieved by prefix delta encoding of the word and error encoding of the pronunciation.

Compact text-to-phone pronunciation dictionary

An utterance detector for speech recognition is described. The detector consists of two components. The first part makes a speech/non-speech decision for each incoming speech frame. The decision is based on a frequency-selective autocorrelation function obtained by speech power spectrum estimation, frequency filter, and inverse Fourier transform. The second component makes utterance detection decision, using a state machine that describes the detection process in terms of the speech/non-speech decision made by the first component.

Automatic utterance detector with high noise immunity

The search network in a speech recognition system is reduced by parsing the incoming speech expanding all active paths ( ), comparing to speech models and scoring the paths and storing recognition level values at the slots ( ) and accumulating the scores and discarding previous slots when a word end is detected creating a word end slot ( ).

Minimization of search network in speech recognition

A memory management method is described for reducing the size of memory required in speech recognition searching. The searching involves parsing the input speech and building a dynamically changing search tree. The basic unit of the search network is a slot. The present invention describes ways of reducing the size of the slot and therefore the size of the required memory. The slot size is reduced by removing the time index, by the model_index and state_index being packed and by a coding for last_time field where one bit represents a slot is available for reuse and a second bit is for backtrace update.

Method of memory management in speech recognition

A method for N-best search for continuous speech recognition with limited storage space includes the steps of Viterbi pruning word level (same word, different time alignment, thus non-output differentiation) states and keeping the N-best sub-optimal paths for sentence level (output differentiation) states.

N-best search for continuous speech recognition using viterbi pruning for non-output differentiation states

Phonetic modeling includes the steps of forming triphone grammars (,) from phonetic data, training triphone models (,), clustering triphones (,) that are acoustically close together and mapping unclustered triphone grammars into a clustered model (,). The clustering process includes using a decision tree based on the acoustic likelihood and allows sub-model clusters in user-definable units.

Method of phonetic modeling using acoustic decision tree

A speed up speech recognition search method is provided wherein the number of HMM states is determined and a microslot is allocated for Hidden Markov Models (HMMs) below a given threshold level of states. A macroslot treats a whole HMM as a basic unit. The lowest level of macroslot is a phone. If the number of states exceeds the threshold level a microslot is allocated for this HMM.

Speed up speech recognition search using macro evaluator

Improved speech recognition is achieved according to the present invention by use of between word and/or between phrase coarticulation. The increase in the number of phonetic models required to model this additional vocabulary is reduced by clustering 19, 20 the inter-word/phrase models and grammar into only a few classes. By using one class for consonant inter-word context and two classes for vowel contexts, the accuracy for Japanese was almost as good as for unclustered models while the number of models was reduced more than half.

Growing community of inventors

Plano, TX, United States of America

Yu-Hung Kao