92 results about "Vector quantisation" patented technology

Vector quantization provides font contour data reduction. Characters are defined in a font collection using segmented outline forms and common references stored to curve segments for similar segments in different characters of the font collection. Compressing the font collection preferably includes applying an entropy encoding data reduction. The common references are stored in a segment storage location. Another aspect includes decompressing a font store containing a compressed collection of outline fonts. The uncompressed curve segments are scaled so as to match a requested size for a letterform.

The invention relates to an embedded, Chinese-English mixed language supporting, embedded application oriented voice recognition method and system for non-specific people. The invention adopts acoustical model trained by mass voice data, acoustical modeling unit set compatible with Chinese and English pronouncing mode, so as to implement Chinese-English mixed voice recognition for non-specific people. According to the invention, a plurality of background models are adopted, a Gauss mixed model (GMM) parameter is obtained by an average adaptive training executed by the background models, then a vector quantization to the difference between the average of the Gauss mixed model and the average of the background models, and the model parameters are compressed. In the recognition stage, rapid Gauss selection, acoustic score pre-calculation, and a simplified GMM model are used, so that the amount of recognition calculation and storage space of the models are greatly reduced, and the voice recognition method and system is applicable on various kinds of embedded application systems.

The invention comprises a method for lossy data compression, akin to vector quantization, in which there is no explicit codebook and no search, i.e. the codebook memory and associated search computation are eliminated. Some memory and computation are still required, but these are dramatically reduced, compared to systems that do not exploit this method. For this reason, both the memory and computation requirements of the method are exponentially smaller than comparable methods that do not exploit the invention. Because there is no explicit codebook to be stored or searched, no such codebook need be generated either. This makes the method well suited to adaptive coding schemes, where the compression system adapts to the statistics of the data presented for processing: both the complexity of the algorithm executed for adaptation, and the amount of data transmitted to synchronize the sender and receiver, are exponentially smaller than comparable existing methods.

The system carries out conversion of digital video signals organized in blocks of pixels from a first format to a second format. The second format is a format compressed via vector quantization. The vector quantization is performed by means of repeated application of a scalar quantizer to the pixels of said blocks with a quantization step (Q) determined in an adaptive way according to the characteristics of sharpness and/or brightness of the pixels and representing said vector quantization in a n-dimensional space indicative of the characteristics on n of said pixels in the block partitioned into cells of size proportional to said quantization step, each cell being assigned to an appropriate binary code, wherein said process further includes identifying at least one symmetry element in said n-dimensional space suitable for separating at least two symmetrical set of cells, and selecting one of said at least two symmetrical set of cells for the assignment of said binary codes. A symmetrical permutation on the n pixels of the block is performed according the selection and a part of said binary code indicating the status of said symmetrical permutation is conveniently set.

Provided is a method for converting a dimension of a vector. The vector dimension conversion method for vector quantization includes the steps of: extracting a specific parameter having a pitch period from an input speech signal and then generating a vector of a dimension that varies according to the pitch period; dividing an entire frequency domain of the generated vector of the variable dimension into at least two frequency domains; and converting the vector of the variable dimension into vectors of mutually different fixed dimensions according to the divided frequency domains. Thereby, not only an error due to the vector dimension conversion is suppressed but codebook memory required for the vector quantization is effectively reduced.

The invention discloses a method refining a parameter of a narrow band vocoder on a decoding end. Based on the coherence of an excitation parameter and a sound track parameter, all parameters are more fine reconstructed on the decoding end, thus the quantization precision is improved, and the quality of synthetic speech sound is further improved. More particularly, for encoding and decoding parameters in a narrow band low-rate speed coding based on a hybrid excitation linear prediction model, a method based on mapping and refining among different parameters is adopted. Various encoding and decoding parameters are quantized by adopting an independent vector through an original technology. According to the invention, the coherence between the excitation parameter and the sound track parameter is considered, and various inverse-quantized encoding and decoding parameters are refined by adopting a nonlinear mapping method, thus the quantization efficiency of various parameters is increased, and the quality of the synthetic speech sound is improved. According to the method, the natural degree of the synthetic speech sound can be improved.

A method and apparatus for reducing the complexity of linear prediction analysis-by-synthesis (LPAS) speech coders. The speech coder includes a multi-tap pitch predictor having various parameters and utilizing an adaptive codebook subdivided into at least a first vector codebook and a second vector codebook. The pitch predictor removes certain redundancies in a subject speech signal and vector quantizes the pitch predictor parameters. Further included is a source excitation (fixed) codebook that indicates pulses in the subject speech signal by deriving corresponding vector values. Serial optimization of the adaptive codebook first and then the fixed codebook produces a low complexity LPAS speech coder of the present invention.

Improved techniques are disclosed for quantizing complex vectors in communication systems. For example, a method includes the following steps. At least one complex vector representative of at least one element of a communication system is obtained. A codeword that approximates the complex vector is identified. The identified codeword is a codeword, from a set of codewords, wherein a real part of a product of the codeword and a scaled version of the complex vector is about maximal over the set of codewords. The scaled version of the complex vector is the product of the complex vector and a constant from a set of constants. In one embodiment, the element of the communication system that the complex vector represents is a channel between a base station and a user terminal in the communication system.

The method samples audio signals and divides up the signals into multiple positive and negative half waves. Sounding/not sounding determination is carried out for half wave based on up/low threshold frequency. Sub code books for corresponding dimension are searched respectively for multiple half waves determined as sounding based on number of sampled points contained between two zero points i.e. vector dimension of the half wave. Vector quantization is carried out for the sub code books to form output index. Since length of sub code book corresponding to each half wave is much smaller than length of total code book, the method reduces operation quantity of searching codes in vector quantization without lowering hearing quality of decoded audio signal. The invention also discloses decoding method and devices for executing the encoding and decoding methods.

The invention discloses an audio coding method and device for optimizing frequency domain noise shaping, and belongs to the technical field of audio coding and decoding. The method comprises the following steps: calculating a voice energy entropy of an audio signal according to the audio signal subjected to low-delay improved discrete cosine transform; judging whether the current frame is a non-voice frame or not according to the voice energy entropy of the current frame of the audio signal and a preset threshold value to obtain a non-voice frame audio signal; and after the non-speech frame audio signal is subjected to first-stage split vector quantization of frequency domain noise shaping, not performing second-stage pyramid vector quantization of a frequency domain noise shaping process. By adding a small amount of operation, the frequency of calling the second-stage pyramid vector quantization by the frequency domain noise shaping module is reduced under the condition that the tone quality is almost unchanged, and the total operand of the frequency domain noise shaping module is reduced.

The invention relates to an engine fault diagnosis method and an engine fault diagnosis device based on exhaust noise vector quantitative analysis. The method and the device mainly solve the problems of complex system and high cost because various parameters are required to be combined to complete engine fault diagnosis in the prior art. The diagnosis device comprises a characteristic extractor, a codebook generator, a memory and a recognizer, wherein the characteristic extractor is respectively connected with the codebook generator and the recognizer, the codebook generator is connected with the memory and the memory is connected with the recognizer. The Mel-frequency cepstrum coefficient (MFCC) of exhaust noise signals are used as a characteristic parameter vector sequence of a signal source in a vector quantitative algorithm, the vector quantitative algorithm is used for analyzing the exhaust noise of an engine, the difference of the exhaust noise of the engine under different working conditions is recognized and therefore the faults of the engine are diagnosed. Not only can nondestructive detection be realized, but also the system cost is low and the signals can be acquired more quickly and effectively.

The invention provides a recognition method for recognizing the on-line handwritten Chinese and Japanese. According to the method, the structured dictionary representation and the vector quantization (VQ) are adopted to establish a compact MRF-based online character recognition method, and the method can be used for recognizing large-scale Chinese and Japanese character sets. Characters are decomposed into the free radicals of basic elements for constituting characters. By means of an MRF-based model, identical free radicals for forming different characters are shared, so that the storage space of a dictionary model is saved. In addition, the method simultaneously applies the VQ technique to compress a character recognizer. Therefore, the storage space of the dictionary model is further greatly compressed on the premise that the recognition rate is not lost at all.

The present invention discloses an iteration optimum design method of selecting predictive vector quantification. The method comprises the steps of dividing training data by using conventional correlation between front and back frames, performing predictive parameter calculation and codebook design by using different training data, quantizing the training data by using designed predictive parameter and codebook and adjusting the training data division according to the quantization error value, and performing optimized selection of predictive vector quantification codebook through the training data division and codebook design iteration. According to the present invention, defect of fixed division of training data is avoided for effectively improving the performance of selecting predictive vector quantification design. The method also reduces the parameter quantizing distortion and improves the signal compression quality.

A feature parameter vector grouping quantization method in speech recognition relates to a vector quantization method in speech recognition methods, belonging to the field of biological recognition. The current feature parameter vector quantization method in the speech recognition generally establishes a code book for multiple words; as speech information differences among the words are greater, and the range of the code book is limited, frames of different words can fall in a same code word region by the quantization of the code book, thus causing problems for improving the subsequent speech recognition rate. The feature parameter vector grouping quantization method leads all speech training information of all words to constitute a speech group, the code book after the vector quantization has great relevance to the information of the corresponding group, feature frames containing the same information in the speech information of different people for the same word can fall in the same code word region with very large probability, thereby providing a basis for the follow-up accurate calculation of various parameters.

The invention discloses a high-dimensional feature processing method and device, and the method comprises the steps: selecting a sample feature, carrying out the coarse quantification of the sample feature, and generating a plurality of clustering centers; respectively adding the high-dimensional features to be inserted into the corresponding HNSW clusters according to the principle of closest distance; and respectively calculating the distances between the target features and the clustering centers, performing HNSW algorithm retrieval in a preset number of clusters closest to the sequence, and returning a retrieval result. In the embodiment of the invention, clustering is carried out by using coarse quantization, so that the method is suitable for large-scale distributed fragmentation storage and is easy for expansion and contraction of mass storage; due to the fact that HNSW algorithm retrieval is carried out in the preset number of clusters closest to the sequence, the recall rate is increased; and, according to the method, the advantages of vector quantization and HNSW are combined at the same time, and the calculation cost and time of data insertion and retrieval are reduced.

The invention provides a video big data accurate retrieval method and system oriented to network live broadcast, and relates to the technical field of computers. Comprising the following steps: carrying out vector quantization coding on audio information in a video to be retrieved and audio signals of each video to be detected in a preset mass database; then, the similarity of the audio signals is calculated through Euclidean distance; carrying out key frame extraction according to each audio similarity result; then super-resolution reconstruction is carried out respectively; carrying out significance detection to generate corresponding core regions; respectively carrying out sparse coding representation; calculating the similarity between a new to-be-retrieved video key frame image and a plurality of new to-be-detected video key frame images; and finally, screening the plurality of videos to be detected to obtain and display a similar video set to the user, so as to realize deep audio similarity detection and video similarity detection, and effectively improve the accuracy of network live broadcast video data retrieval.