Application of on-line filler modeling in specific domain speech recognition system
在线废料建模在特定领域语音识别中的应用

It is difficult for a speech recognition system in specific domain to deal with parentheses or sentence particles which is not depicted in rule since it instructs acoustic layer recognition in strict syntax rules model.To solve this problem,an on-line filler modeling method based on wordgraph decoder was applied to specific domain speech recognition system.The choice of the parameter N in this method was discussed in detail,and the correlation between materials' SNR and the parameter N was also analyzed and validated.Moreover,a model parameter optimization method,which benefits from the correlation,was proposed.Computer simulation validates that the proposed method increases recognition rate by 18.34% for sentence and 11.47% for slot relatively.