Design and collecting the speaker independent data set of voice commands for controlling the smart home appliances based on Persian speech

EasyChair Preprint 1641

Abstract

Nowdays, whit advancement of technology and speech processing systems, in addition to human to human interactions, speech is also used  in human to machine interaction. One aspect of using speech in areas such as smart homes is to replace touch commands with voice-based commands. For this purpose, a voice command recognition system can be designed and implemented. The system has the capability to easily realize users needs using simple spoken commands. Due of unavailability of Persian voice commands for controlling smart home appliances, this paper studies the process of designing, assembling and evaluating dataset of speaker independent voice commands for controlling smart home applications (TV, voice recorder, lamp) base on Persian speech.  This dataset is divided into two sections: train and test sets. Hidden Markov Model (HMM) and long short time memory (LSTM) have been used to evaluate the dataset. The result of evaluating the data set based on HMM  shows that the trained word-based speech recognizer has the accuracy about 92%,correctnees about 93% and word error rate of only 8%. In addition, the result of the evaluating base on LSTM shows that the accuracy of word-based recognizer with the best configuration is about 98%.

Keyphrases: Hidden Markov Model, Long Short-Term Memory, Persian spoken dataset, Smart Home Appliances, voice commands

@booklet{EasyChair:1641,
  author    = {Leila Safarpoor Kalkhoran and Shima Tabibian and Elaheh Homayounvala},
  title     = {Design and collecting the speaker independent data set of voice commands for controlling the smart home appliances based on Persian speech},
  howpublished = {EasyChair Preprint 1641},
  year      = {EasyChair, 2019}}