| Title |
Skaičių ir vardų garsynų atpažinimo, naudojant fonemomis grįstus paslėptuosius Markovo modelius, tyrimas / |
| Translation of Title |
Investigation of digits and names corpora recognition by phones based hidden Markov models. |
| Authors |
Račys, Remigijus |
| Full Text |
|
| Pages |
55 |
| Keywords [eng] |
corpus ; HTK ; hidden Markov models ; annotation |
| Abstract [eng] |
In this paper the detection experiments of separate commands (digits and names) are presented, in the field of speech recognition, adaptation–phonemes–based hidden Markov models. Experimental research of the corpus level grammatical phonemes are prepared. For a part of the corpus annotations were created “manually”. The research recognition accuracy depends on the used set of phonemes and the use of HTK package tools. Experimenting with various phoneme sets was made in order to find their impact on the detection accuracy. Examination of names and numbers of the isolated command recognition accuracy were 95.36%. This was achieved by using the expanded SAMPA–LT phoneme set. In order to check the influence of recognition annotated corpus, a number of tests were made to assess the accuracy of the number corpus. An annotated corpus for 6 announcers was made in the training stage. Testing was made in the recognition inspection stage while using 24 announcers. With a minimum set of phonemes, the accuracy obtained – 99.13%, and using SAMPA–LT phoneme set the result obtained was 98%. Contextual phonemes (triphones) were created with which the detection accuracy of the minimum set was 99.10%, while the recognition accuracy using triphones created from SAMPA–LT set – 98%. |
| Dissertation Institution |
Kauno technologijos universitetas. |
| Type |
Master thesis |
| Language |
Lithuanian |
| Publication date |
2017 |
