Verification Mode Research Articles

A talker recognition based on vector quantization codebooks

Highly efficient and effective short‐term spectral representations of talkers can be obtained using vector quantization (VQ) codebook construction techniques. A talker recognition system has been implemented in which each talker is represented by a VQ codebook constructed from a large set of short‐term spectral vectors obtained from a series of training utterances provided by the talker. In operation, the utterances of an unknown talker are analyzed and “encoded” using the codebook of a specified talker. The accumulated distortion between the input utterances and the specified talker's codebook is used to carry out a talker recognition decision. This technique can be said to be text independent to the extent that the training utterances adequately represent each talker's speech sound repertoire. The system can be extended to text‐dependent operation with an additional training procedure in which specified utterances provided by a given talker are represented as encoded vector sequences using the talker's codebook. In use, an unknown talker is prompted to provide specified utterances which are analyzed and compared with the encoded prototypes for a specified talker. The system has been evaluated using a 100‐talker database of 20 000 digits spoken in isolation. In a talker verification mode, average equal‐error rate performance of 2.2% for text‐independent operation and 0.3% for text‐dependent operation is obtained for seven‐digit‐long test utterances.

Talker recognition in tandem with talker-independent isolated word recognition

A talker recognition system operating in tandem with a talker-independent isolated word recognizer is described and evaluated. The word recognizer uses a small set of reference templates for each vocabulary word. Each set is intended to span and typify individual talker templates over a large population of talkers. Word recognition decisions are based on template distance scores obtained by comparing processed input utterances to each set of reference templates. The distribution of distance scores for the templates corresponding to the actual word input has been found to be reasonably consistent for individual talkers, and to vary sufficiently from talker to talker to provide the basis for a talker recognition capability. A system has been implemented to exploit this capability. An evaluation of the system, carried out using a 100-talker database of digit utterances, shows that good talker recognition performance can be obtained for input utterances consisting of sequences of seven or more digits. Identification error rates varying from 3.6 to 14.0 percent for talker populations varying from 10 to 100 talkers are obtained. When the recognizer orders the talkers as candidates for recognition, the correct talker is found, on the average, among the top 0.8 percent of the population. Tested in a talker verification mode, the average error rate is approximately 8 percent.