Loading...

A Semi Automatic System for Procurement Continues Speech Corpus with Using Human Computing

Ramezani, Ali Akbar | 2010

483 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 41369 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Abolhassani, Hassan
  7. Abstract:
  8. Semi-automated Continuous Speech system deals with collecting corpus speech. It has very different problems in various natural languages because of different types of speech, and collecting speech corpus containing continuous word is very time-and- cost consuming. To overcome this problem, human-based computing techniques are proposed to collect speech corpus. In this thesis, we describe influencing factors in collecting steps of continuous speech and provide a description of human-based techniques to solve problems specially collecting of continuous speech. To collect speech corpus using human-based techniques, a game has been used so that speech corpus of web users are collected during the game. To collect sentences related to the continuous speech of Farsi, we have used subtitles of dubbed movies which has been proved that by growing text corpus to infinite, collected corpus will be convergent in phonetic balancing point of view. To collect new corpus, 90 men and women speaker in 18 to 25 and 35 to 40 age groups and over 2160 sentences have been used so that after refining them 1900 voice which representative of 1900 sentences are studied for training and test. The result of experiments showed that for particular applications for which the cost and time of collecting speech corpus is much more important than effectiveness of learning, and also because of collecting type (method),diversity of voice recording environments and utilized microphones which are very similar to web data ( applications like search in search engines which operate based on speech and have speech information retrieval in web) using this technique for data collecting is effective and making use of these corpus has better performance. The performance of new collected corpus is about 40% which has been decreased about 20% comparing Farsdot, but the advantage of this corpus over Farsdot is that the cost and time for collecting speech corpus has been decreased and also it has less steep drop in performance for noisy speech compared to Farsidot and consequently it is more resistant to noise
  9. Keywords:
  10. Human Base Computation ; Speech Corpus ; Continuous Speech ; Semi Automatic System

 Digital Object List

 Bookmark

No TOC