HMM-Based Recognition and Adaptation of Persian Children's Speech

There are high variability in children's speech compared to adults' which is mainly because of their shorter vocal tract length and smaller vocal fold which results in lower accuracy in speech recognition task (about 54.5% in this work). Therefore using adaptation techinques which reduce these variabilities has been suggested. In this paper we focused on the problem of speech recognition for Persian children using adaptation techniques performed on two models: one trained on children's speech and one on adults'. We used a speaker normalization method which is combination of vocal tract length normalization and model adaptation. Experiments have shown that using adult model has low performance which inreases 37% when using adaptation techniques. It is also shown that using these 222 G. Tadayon Tabrizi, S. Setayeshi and M. Molavi Kakhki techniques will increase recognition rate by 7% when using recognizer trained on children's speech.