Data processing: speech signal processing – linguistics – language – Speech signal processing – Recognition
Reexamination Certificate
2000-03-30
2001-10-16
Dorvil, Richemond (Department: 2741)
Data processing: speech signal processing, linguistics, language
Speech signal processing
Recognition
C704S251000, C704S231000, C704S255000
Reexamination Certificate
active
06304844
ABSTRACT:
FIELD OF THE INVENTION
This invention relates generally to speech recognition systems and more specifically to a speech recognition system for mobile communication devices.
BACKGROUND OF THE INVENTION
Transmission of information from humans to machines has been traditionally achieved though manually-operated keyboards, which presupposes machines having dimensions at least as large as the comfortable finger-spread of two human hands. With the advent of electronic devices requiring information input but which are smaller than traditional personal computers, the information input began to take other forms, such as pen pointing, touchpads, and voice commands. The information capable of being transmitted by pen-pointing and touchpads is limited by the display capabilities of the device (such as personal digital assistants (PDAs) and mobile phones). Therefore, significant research effort has been devoted to speech recognition systems for electronic devices. Among the approaches to speech recognition by machine is for the machine to attempt to decode a speech signal waveform based on the observed acoustical features of the signal and the known relation between acoustic features and phonetic sounds. This acoustic-phonetic approach has been the subject of research for almost 50 years, but has not resulted in much success in practice (rf.
Fundamentals of Speech Recognition
, L. Rabiner & B. H. Juang, Prentice-Hall). Problems abound, for example, it is known in the speech recognition art that even in a speech waveform plot, “it is often difficult to distinguish a weak, unvoiced sound (like “f” or “th”) from silence, or a weak, voiced sound (like “v” or “m”) from unvoiced sounds or even silence” and there are large variations depending on the identity of the closely-neighboring phonetic units, the so-called coarticulation of sounds (ibid.). After the decoding, the determination of the word in the acoustic-phonetic approach is attempted by use of the so-called phoneme lattice which represents a sequential set of phonemes that are likely matches to spoken input. The vertical position of a phoneme in the lattice is a measure of the goodness of the acoustic match to phonetic unit (“lexical access”). But “the real problem with the acoustic-phonetic approach to speech recognition is the difficulty in getting a reliable phoneme lattice for the lexical access stage” (ibid.); that is, it is almost impossible to label an utterance accurately because of the large variations inherent in any language.
In the pattern-recognition approach, a knowledge base of versions of a given speech pattern is assembled (“training”), and recognition is achieved through comparison of the input speech pattern with the speech patterns in the knowledge base to determine the best match. The paradigm has four steps: (1) feature extraction using spectral analysis, (2) pattern training to produce reference patterns for an utterance class, (3) pattern classification to compare unknown test patterns with the class reference pattern by measuring the spectral “distance” between two well-defined spectral vectors and aligning the time to compensate for the different rates of speaking of the two patterns (dynamic time warping, DTW), and (4) decision logic whereby similarity scores are utilized to select the best match. Pattern recognition requires heavy computation, particularly for steps (2) and (3) and pattern recognition for large numbers of sound classes often becomes prohibitive.
Therefore, systems relying on the human voice for information input, because of the inherent vagaries of speech (including homophones, word similarity, accent, sound level, syllabic emphasis, speech pattern, background noise, and so on), require considerable signal processing power and large look-up table databases in order to attain even minimal levels of accuracy. Mainframe computers and high-end workstations are beginning to approach acceptable levels of voice recognition, but even with the memory and computational power available in present personal computers (PCs), speech recognition for those machines is so far largely limited to given sets of specific voice commands. For devices with far less memory and processing power than PCs, such as PDAs, mobile phones, toys, entertainment devices, accurate recognition of natural speech has been hitherto impossible. For example, a typical voice-activated cellular phone allows preprogramming by reciting a name and then entering an associated number. When the user subsequently recites the name, a microprocessor in the cell phone will attempt to match the recited name's voice pattern with the stored number. As anyone who has used present day voice-activated cell phones knows, the match is sometimes inaccurate (due to inconsistent pronunciation, background noise, and inherent limitations due to lack of processing power) and only about 25 stored numbers are possible. In PDA devices, it is necessary for device manufacturers to perform extensive redesign to achieve even very limited voice recognition (for example, present PDAs cannot search a database in response to voice input).
As for spelling words for voice input, there is the problem with the confusable sets: {A,J,K}, {B,C,D,E,G,P,T,V,Z}, {Q,U}, {I,Y}, and {F,S,X}. These can generally only be discriminated based upon a small, critical portion of the utterance. Since conventional recognition relies on a simple accumulated distortion score over the entire utterance duration (a binary “yes” or “no”), this does not place sufficient emphasis on the critical parts resulting in poor recognition accuracy. Clearly, an approach would be to weight the critical portions, but this method has not achieved high recognition accuracy and carries a heavy computational burden.
In summary, the memory and computation necessary for accurate and fast voice recognition also require increased electrical power and complex operating systems; all of these carry increased cost. Thus present voice recognition technology is not feasible for mobile communication devices because of their weight, electrical power requirement, complexity, and cost.
SUMMARY OF THE INVENTION
There is a need, therefore, for an accurate speech recognition system capable of rapidly processing greater varieties of words and operable in many different devices, but without the computational power and memory requirements, high power consumption, complex operating system, high costs, and weight of traditional systems so that voice information transfer is feasible for both person-to-person and person-to-machine communication for mobile phones, PDAs, electronic toys, entertainment products, and any other devices requiring communication. This invention provides accurate speech recognition for electronic devices with low processing power and limited memory storage capability. Accuracy is achieved primarily through the utilization of individual alphanumeric character utterances to transmit words, thereby overcoming the lexical access problem. Because individual alphanumeric utterances produce speech waveforms which change very little over the time region and are separated by a pause (the typical pause between dictating individual characters or letters), this invention solves the problem of distinguishing weak, unvoiced sounds and the indefiniteness associated with coarticulation. This invention comprises a microphone, front-end signal processor for generating parametric representations of speech input signals, a pronunciation database, a letter similarity comparator for comparing the parametric representation of the input signals with the parametric representations of letter pronunciations, and generating a sequence of associations between the input speech and the letters in the pronunciation database, a vocabulary database, a word comparator for comparing an aggregated plurality of the letters with the words in the vocabulary database and generating a sequence of associations between them, and a display for displaying the selected letters and words for confirmation.
Chang Josephine
Chen Juinn-Yan
Kim Yoon
Pan James
Baker & McKenzie
Chen Robert H.
Dorvil Richemond
Konski Antoinette F.
Nolan Daniel A.
LandOfFree
Spelling speech recognition apparatus and method for... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Spelling speech recognition apparatus and method for..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Spelling speech recognition apparatus and method for... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2567562