Abstract: A recognition system (10) incorporates a filterbank analyser (16) producing successive data vectors of energy values for twenty-six frequency intervals in a speech signal. A unit (18) compensates for spectral distortion in each vector. Compensated vectors undergo a transformation into feature vectors with twelve dimensions and are matched with hidden Markov model states in a computer (24). Each matched model state has a mean value which is an estimate of the speech feature vector. A match inverter (28) produces an estimate of the speech data vector in frequency space by a pseudo-inverse transformation. It includes information which will be lost in a later transformation to frequency space. The estimated data vector is compared with its associated speech signal data vector, and infinite impulse response filters (44) average their difference with others. Averaged difference vectors so produced are used by the unit (18) in compensation of speech signal data vectors.

Abstract: A speech analysis system 10 incorporates a filterbank analyser 18 producing successive frequency data vectors for a speech signal from two speakers. From each data vector, units 22A and 22B produce a set of modified data vectors compensated for differing forms of distortion associated with respective speakers. A computer 24 matches modified data vectors to hidden Markov model states. It identifies the modified data vector in each set exhibiting greatest matching probability, the model state matched therewith, the form of distortion with which it is associated and the model class, ie speech or noise. The matched model state has a mean value providing an estimate of its associated data vector. The estimate is compared with its associated data vector, and their difference is averaged with others associated with a like form of distortion in an infinite impulse response filter bank 48A or 48B to provide compensation for that form of distortion.

Abstract: A speech training aid compares a trainee's speech with models of speech, stored as sub-acoustic word models, and a general speech model to give an indication of whether or not the trainee has spoken correctly. An indication of how well the word has been pronounced may also be given. An adult operator enters the word to be tested into the training aid which then forms a model of that word from the stored sub-word speech models. The stored acoustic models are formed by first recording a plurality of words by a plurality of trainees from a given list of single words. These recordings are then processed off-line to give a basic acoustic model of an acceptable or correct sound for each phoneme in the context of the pre-and proceeding phonemes. The acoustic models are Hidden Markov Models. The limits of acceptable pronunciation, applied to different words and trainees, may be adjusted by variable penalty values applied in association with the general speech acoustic model.

Abstract: A children's speech training aid compares a child's speech with models of speech, stored as sub-word acoustic models, and a general speech model to give an indication of whether or not the child has spoken correctly. An indication of how well the word has been pronounced may also be given. An adult operator enters the word to be tested into the training aid which then forms a model of that word from the stored sub-word speech models. The stored acoustic models are formed by first recording a plurality of words by a plurality of children from a given list of single words. These recordings are then processed off-line to give a basic acoustic model of an acceptable or correct sound for each phoneme in the context of the pre- and proceeding phonemes. The acoustic models are Hidden Markov Models. The limits of acceptable pronunciation applied to different words and children may be adjusted by variable penalty values applied in association with the general speech acoustic model.

Abstract: Semiconductor crystalline materials, e.g. silicon, GaAs, are grown from a melt, e.g. using the Czochralski technique where a seed crystal is dipped into the melt then slowly withdrawn. Rotation of the growing crystal (6) is partly responsible for convective flows within the melt (5). Convective flows are reduced while radial uniformity is improved by subjecting the crystal/melt interface to a shaped magnetic field. This magnetic field is rotationally symmetrical about the axis of crystal rotation, with a component of field parallel to this axis that is less than 500 gauss, preferably less than 200 gauss, with a value above 500 gauss at other parts of the melt. The field may be produced by two superconducting magnet coils (21, 22) spaced apart and arranged co-axially with the axis of crystal rotation.