Use this URL to cite or link to this record in EThOS:
Title: Adaptive speech recognition using vector quantization
Author: Goatcher, J. K.
Awarding Body: University College of Swansea
Current Institution: Swansea University
Date of Award: 1986
Availability of Full Text:
Access from EThOS:
Adaptation to the voice characteristics of different people is performed automatically by human listeners. If a machine is to achieve the performance approaching that of a human then it too must adapt to the individual talker. One such adaptation scheme is proposed and investigated in this thesis. A review of template-based speech recognition systems is presented. The spectral estimation technique of LPC and vector quantization are examined in some detail. The development of a speech recognition system is described including performance measurements and comparisons with other published results. The effect of vector quantization upon the system's performance is presented for a range of vector codebook sizes from 2 to 128. The hardware for the generation of a database of speech utterances is described. This database is used for all recognition and adaptation experiments and consists of the numeric and alpha-numeric vocabulary sets along with two phonetically representative sentences. The proposed adaptation scheme is based upon the assumption that a mapping may be made from the codebook entries of speaker-independent templates to the codebook entries of a particular speaker. A conventional dynamic time warping algorithm gives a trace of the correspondence between the codebook entries of a particular speaker and the speaker-independent templates. A map is created with the scores of this matching process and characterises the differences. These maps improve the performance of the speaker-independent templates, an error rate of 7.5% is reduced to 2.2% for the numeric vocabulary and from 28.4% to 19.2% for the alpha-numeric vocabulary. A number of variations on the method for generating the maps are presented. The most succesful of which is the simplest, using the whole vocabulary to form the maps.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available