U.S. Philips Corporation

Hermann Ney

Michael Harald Kuhn

Volker Steinbiss

Reinhold Hab-Umbach

Horst H Tomaschewski

Rudolf Geppert

Reinhard Kneser

Herbert Piotrowski

Manfred Tasto

Dieter Mergel

Bach-Hiep Tran

Stephan Vogel

Christoph Tillmann

Andreas Noll

Lutz Welling

For translating a word-organized source text into a word-organized target text through mapping of source words on target words, both a translation model and a language model are used. In particular, alignment probabilities are ascertained between various source word & target word pairs, whilst preemptively assuming that alignment between such word pairs is monotonous through at least substantial substrings of a particular sentence. This is done by evaluating incrementally statistical translation performance of various target word strings, deciding on an optimum target word string, and outputting the latter.

Method and device for translating a source text into a target using modeling and dynamic programming

In a frequently used speech synthesis for voice output an excitation signal is applied to a number of resonators whose frequency and amplitude are adjusted in accordance with the sound to be produced. These parameters for adjusting the resonators may be gained from natural speech signals. Such parameters gained from natural speech signals may also be used for speech recognition, in which these parameter values are compared with comparison values. According to the invention, the parameters, particularly the formant frequencies, are determined by forming the power density spectrum via discrete frequencies from which autocorrelation coefficients are formed for consecutive frequency segments of the power density spectrum from which, in turn, error values are formed, while the sum of the error values is minimized over all segments and the optimum boundary frequencies of the segments are determined for this minimum. Via the autocorrelation coefficients, the LPC predictor coefficients can then be computed, from which coefficients the formant frequency is computed. The minimum of the error sum for the individual segments is found by way of dynamic programming, in which auxiliary values are initially computed from the power density spectrum and stored as Tables from which the autocorrelation coefficients are easily determined for individual frequency segments which are required for the computations in the dynamic programming process.

Method of deriving characteristics values from a speech signal

A method and apparatus for processing a sequence of words in a speech signal for speech recognition. The method includes the steps of sampling, at recurrent instants, said speech signal for generating a series of test signals. Signal-by-signal matching and scoring is generated between the test signals and a series of reference signals, where each of the series of reference signals forms one of a plurality of vocabulary words arranged as a vocabulary tree. The vocabulary tree includes a root and a plurality of tree branches wherein any tree branch has a predetermined number of reference signals and is assigned to a speech element and any vocabulary word is assigned to a particular branch junction or branch end. Acoustic recombination determines both continuations of branches and the most probable partial hypotheses within a word because of the use of a vocabulary built up as a tree with branches having reference signals. At least one complete word for a particular test signal is determined, and, separately, for each completed word, there is: I) a word result formed including a word score and an aggregate score, said aggregate score derived from said word score and from a language model value assigned to a combination of said completed word and a uniform-length string of prior completed words.

Method and apparatus for recognizing spoken words in a speech signal by

In the recognition of coherently spoken words, a plurality of hypotheses is usually built up which end in various words during the recognition process and are then to be continued with further words. To keep the number of words yet to be continued as small as possible, especially in the case of a large vocabulary, it is known to carry out a look-ahead in a limited time space. It is suggested according to the invention to use the same phonemes for the look-ahead as for the actual recognition and to add together the differential sums obtained in the look-ahead for the evaluation of the partial hypothesis which has just ended and which is to be continued, and to compare this sum with a threshold value which depends on the extrapolated minimum total evaluation at the end of the time space of the look-ahead. The searching space for hypotheses to be continued can be limited by this in a particularly favorable manner.

Speech recognition apparatus and method using look-ahead scoring

When a language model is to be used for the recognition of a speech signal and the vocabulary is composed as a tree, the language model value cannot be taken into account before the word end. Customarily, after each word end the comparison with a tree root is started anew, be it with a score which has been increased by the language model value so that the threshold value for the scores at which hypotheses are terminated must be high and hence many, even unattractive hypotheses remain active for a prolonged period of time. In order to avoid this, in accordance with the invention a correction value is added to the score for at least a part of the nodes of the vocabulary tree; the sum of the correction values on the path to a word then may not be greater than the language model value for the relevant word. As a result, for each test signal the scores of all hypotheses are of a comparable order of magnitude. When a word end is reached, the sum of all added correction values is subtracted and the correct language model value is added.

Method of recognizing a sequence of words and device for carrying out

For the recognition of coherently spoken speech with a large vocabulary, language model values which take into account the probability of word sequences are considered at word transitions. Prior to the recognition, these language model values are derived on the basis of training speech signals. If the amount of training data is kept within sensible limits, not all word sequences will actually occur, so that the language model values for, for example an N-gram language model must be determined from word sequences of N-1 words actually occurring. In accordance with the invention, these reduced word sequences from each different, complete word sequence are counted only once, irrespective of the actual frequency of occurrence of the complete word sequence or only reduced training sequences which occur exactly once in the training data are taken into account.

Single-count backing-off method of determining N-gram language model

In the recognition of coherent speech, language models are favourably used to increase the reliability of recognition, which models, for example, take into account the probabilities of word combinations, especially of word pairs. For this purpose, a language model value corresponding to this probability is added at boundaries between words. In several recognition methods, for example, when the vocabulary is built up from phonemes in the shape of a tree, it is not known at the start of the continuation of a hypothesis after a word end which word will actually follow, so that a language model value cannot be taken into account until at the end of the next word. Measures are given for achieving this in such a manner that as far as possible the optimal preceding word or the optimal preceding word sequence is taken into account for the language model value without the necessity of constructing a copy of the searching tree for each and every simultaneously ending preceding word sequence.

Method and apparatus for recognizing spoken words in a speech signal

During the recognition, speech values which are derived from sample values of the speech signals are compared with reference values, the words of a given vocabulary each time being given by a sequence of reference values. The words are then determined from phonemes according to a fixed pronouncing lexicon and the reference values for the phonemes are determined in a learning phase, each phoneme within a word consisting of a number of equal reference values determined in the learning phase. In order to approach transitions between phonemes, each phoneme may also consist of three sections of each time constant reference values. By the given number of reference values per phoneme, the time duration of a phoneme in a given word can be simulated more accurately. Different possibilities are indicated to determine the reference values and the distance value during the recognition.

Method of recognizing coherently spoken words

A method of recognizing continuously spoken words in which, during the speech recognition, speech values derived from the speech signal are compared with comparison values of the individual words of a given vocabulary. In order to reduce the rate of recognition errors, it is essentially known to take into account speech models, which consequently admit only selected sequences of words. From the theory of the formal languages, a class of speech models designated as 'context-free grammar' is known, which represents a comparatively flexible speech model. For the use of this speech model in the technical process of recognition of a speech signal two lists are now utilized, which indicate the assignment between words and given syntactical classes and the assignment of these classes to, as the case may be, two other classes. Both lists are used at each new speech signal in that there is constantly considered backwards, which class explains most clearly the preceding speech section. At the end of the speech signal, starting from the class indicating the whole sentence, the sequence of the words can be followed backwards, which has yielded the smallest total distance sum and which moreover fits into the speech model given by the two lists.

Method of recognizing continuously spoken words

In a method of and an arrangment for determining the start-point and end-point of a word signal in a speech signal consisting of isolated utterances, three adjacent windows are determined at each new digital value for the last arrived stored digital values, in which the central window contains the actual word signal. The length of this central window is varied for each digital value between a minimum and a maximum value, and a threshold value is determined from the two adjacent windows and is subtracted from the energy contained in the central window. Thus, the method and the apparatus always takes the overall speech signal into account instead of individual isolated portions so that a reliable end-point determination then is possible.

Method of and apparatus for determining start-point and end-point of

In a speech or speaker recognition system, a segment or sequence of speech parameter values are smoothed to a most probable sequency by Dynamic Programming. The method for determining the variation with time of a speech parameter is based on a speech signal which is subdivided into successive segments and an individual value exists in each segment and for each value of the parameter within a limited range of values. For the example of the fundamental voice frequency, a value has been generated in each speech segment with the aid of the AMDF (Average Magnitude Difference Function). The required variation now links a sequency of horizontally, vertically or diagonally directly adjacent speech parameter values to one another in such a manner that the sum of the associated individual values represents a minimum. In this arrangement, this sum is slightly magnified in diagonal or vertical sections since a horizontal variation is most probable. This magnification is controlled by certain fixed values which influence the smoothness of the variation.

Method for determining the variation with time of a speech parameter and

The degree of mis-match which would be obtained between a test and a reference signal, for example speech signals, should their time-axes be subjected to that relative shift and/or distortion which is required to minimize the degree of mismatch is carried out by sampling the two signals at regular intervals and storing these samples in memories. All the samples of one signal are then read out in succession from one memory; each successive sample of the other signal is read out from the other memory, and the difference between each pair of samples is formed in a subtractor. Each difference value from the subtractor is added to the smallest of three quantities, X, Y and Z and the result stored in a register to form the new quantity X. This new quantity X is also stored in a further memory cyclically addressed in tandem with the memory 12. The quantity Z is the previous content of the location of memory into which the new quantity X is written, and the quantity Y is the previous value of the quantity Z, obtained by shifting the previous value of the quantity Z in a two-stage shift register. Thus the quantities X, Y and Z correspond to the present sample of the other signal and the immediately preceding sample of the one signal, the immediately preceding sample of both signals, and the present sample of the one signal and the immediately preceding sample of the other signal respectively. The result is that each successive value of the quantity X is the optimum cumulative distance value between the two signals for the corresponding pair of samples should these be assigned to each other when matching the two signals to each other. The process is continued until all the sampled value have been read and paired, at which point the value of the quantity X is a measure of the above-defined degree of mis-match.

Measuring mis-match between signals

In a device for adjusting a movable electro-acoustic sound transducer, an optical duct is connected to the sound transducer which only allows an object, such as a display table or a data display unit, to be viewed completely in one specific position. This position is unambiguously reproducible and so is the position of the sound transducer or microphone relative to the mouth of the speaker. The optical duct may comprise diaphragms or a tube, possibly with intermediate walls, or a phase amplitude grating. The object may be a display unit and the device further comprises a device for generating variable data on the display. If used for speaker identification, the variable data generator projects consecutive instructions on the display so that, the speech recognition process is controlled automatically. The speaker to be identified is then required to keep his mouth in the correct position relative to the microphone.

Device for adjusting a movable electro-acoustic sound transducer

In order to reduce the information contents of a sound signal, for example a speech signal, the sound signal is divided into adjacent frequency ranges and the signals then obtained are consecutively integrated over short time intervals to short-time-spectrum signals. Subsequently, a histogram of the short-time-spectrum signals is formed for each frequency range and said histograms are divided into a given number of equal-area portions. The boundaries of said portions, which are referred to as quantiles, represent the characteristic values. The amplifier noise or the noise of the complete recording arrangement as well as a difference in sound energy level during when different sound signals are processed can then be eliminated at least in respect of their effect.

Growing community of inventors

Aachen, Germany

Hermann Ney