VSE knjižnice (vzajemna bibliografsko-kataložna baza podatkov COBIB.SI)
  • Razpoznavanje govora preko telefona s polzveznimi Prikritimi modeli Markova
    Imperl, Bojan ; Kačič, Zdravko ; Köhler, Joachim
    In the speech recognition systems that are based on continuous mixture density HMM (CDHMM), the emission probability computation is usually the most complex task. Many methods of fast log-likelihood ... computation have already been proposed where significant acceleration has been reported. In order to achieve further improvements of the HMM performance we have implemented the semi-continuous HMM (SCHMM). The SCHMM are known to be computationally less complex than CDHMM but are usually considered to achieve lower recognition rates unless for the cases of insufficient training data. However, in our experiments described in this paper we have showed that using the proper training procedure the SCHMM recogniser can achive recognition rate as high as CDHMM also for the cases of sufficient amount of training data. The comparative tests of CDHMM and SCHMM were performed on the task of speaker independent isolated digits recognition over the telephone using the Voice - Mail and SNABI databases. THe SCHMM recogniser achieved the same recognition rate as the CDHMM recogniser and the SCHMM recognition runtime was less than 50% of the CDHMM recognition runtime.
    Vir: Zbornik šeste Elektrotehniške in računalniške konference ERK '97, 25. - 27. september 1997, Portorož, Slovenija (Str. B/201-204)
    Vrsta gradiva - prispevek na konferenci
    Leto - 1997
    Jezik - slovenski
    COBISS.SI-ID - 3098390

Vnos na polico