The Consistency Checking System Of Prosodic Structure Marker

Speech synthesis technology is also called text-to-speech(TTS)technology.A major issue speech synthesis technology addresses is the conversion of textual information into acoustic information,namely the humanization of machines in terms of speech-Concerning theoretical research in acoustics,linguistics,digital signal processing,computer science and more disciplinary domains.Naturalness and intelligibility are two key indices of,and criteria on the quality of,speech synthesis.Currently,intelligibility has achieved a high level,while naturalness remains inadequately high.A crucial restraint to naturalness of speech synthesis is computer's incapacity to effectively recognize and simulate the rhythms in natural flow of speech.Over the recent years,scholars' studies on rhythm are mostly based on large-sized corpus.Since corpus is subject to human errors resulting from artificial marking and too long duration of marking,the quality of corpus awaits improving further.Inconsistency in marking is a key aspect restricting the quality of corpus.This article mainly addresses the issue of inconsistency in rhythmic structure marking,in the hope of improving the deficiency in artificially marking corpus through research in this area,thereby making certain contributions to enhancing the naturalness of speech synthesis.This article from the three aspects of the prosodic structure Consistency marked proofread in-depth study:(1)Rule-based prosodic word marked inconsistency discovery and modificationThe rule-based words and rhythm of the original corpus prosodic word comparison,find and display inconsistent,suggesting proofreader or modify reservations.(2)Based on maximum entropy and boundary conditions with the prosodic phrase prediction model to build the airportProsodic structure based on artificial tagging corpus,respectively using the maximum entropy method,CRFs constructed a prosodic phrase prediction model.Atomic characteristic model uses the word,part of speech,word length,distance,etc.and combinations thereof feature construction feature templates and generate the corresponding training model,automatically predict prosodic boundary.(3)Based on multi-policy support prosodic phrase inconsistency found and proofreadingFirst part of speech composition way to find inconsistencies exist in the corpus,as amended,combined with maximum entropy model,prosodic boundary conditions with the airport model predictions found prosodic phrase marked inconsistencies.