基于多特征CNN与DTW的藏语发音相似度计算方法

打开文本图片集

中图分类号：TP311.1 文献标志码：A 文章编号：2095-2945（2025）29-0010-0

Abstract：Speechsimilarityevaluation technologyhasbecomea hot spot inspeech information processing research today. Thereisarelativelylackofresearchonspeech similarityevaluationofTibetanandotherminoritylanguagesinChina.Basedon this，thispaperdiscusesusingaone-dimensionalconvolutionalneuralnetwork（1D-CNN）hybridmodeloffeaturecompresion anddynamictimewarping（DTW）alignmenttodesignamulti-featurefusionframework，andusescontrastlossandanti-noise trainingtosolvetheproblemsoffeatureredundancy，temporaldeformationrobustnessandlowresourcegeneralizationinTibetan pronunciationsimilaritycalculation，andfinallyrealizetheevaluationofpronunciationsimilarityofcommonlyusedTibetanwords. Experimentsonaself-builtdatasetcovering three majordialectsshow that theoverallaccuracyofthemodelreaches 93.2% ， verifying the effectiveness of this method.

Keywords：speechsimilarityevaluation;one-dimensionalconvolutionalneuralnetwork（1D-CNN）;dynamictimewarping; multi-feature fusion; Tibetan pronunciation

本文提出了一种基于1D-CNN特征压缩与DTW对齐的混合模型来计算藏语发音相似度，不仅能够有效应对藏语发音相似度计算中的特征冗余性、时间形变鲁棒性，还可很好地适应低资源泛化性等难题[1]。（剩余4047字）