| The advancement of artificial intelligence technology has been ongoing and significant in recent years.Intelligent speech technology has become an indispensable part of today’s modern society.However,due to language differences between different ethnic groups and regions,cross-lingual barrier-free communication remains an important challenge.Therefore,cross-language conversion technology is receiving increasing attention and research in the field of artificial intelligence.Cross-language conversion refers to the process of using a machine to convert a source language into another with a machine without changing semantics.This technology not only enhances people’s living standards,but also helps promote exchanges between different cultures.At present,there are relatively few research on Chinese-Tibetan cross-language conversion.To this end,this article is aimed at the transformation of Chinese-Tibetan cross-language conversion of speech to text,focusing on Tibetan speech recognition technology and Chinese-Tibetan machine translation techniques,which aims to improve the overall performance of Chinese-Tibetan cross-language conversion.This thesis focuses on the following research content and innovative points:1.Established a recognition corpus based on Tibetan language and a parallel corpus of Chinese-Tibetan bilingual languages.According to Tibetan linguistics knowledge,the corpus is used to build a corpus in the units of vowels and establish a pronunciation dictionary.A total of 18,000 Tibetan corpus was established in Tibetan language recognition corpus,including 20 speakers,16 girls and 4 boys.The Chinese Tibetan bilingual parallel language corpus has a total of 12,000 sentences in parallel,which contains 1000 test corpus and 1000 sentence verification corpus.2.An acoustic model combining CNN and DFSMN was used to study the Tibetan speech recognition technology.In this thesis,the convolutional neural networks(Convolution Neural Network,CNN)learns about the local frequency domain and time domain features,Deep feedforward sequence memory network(Deep Feedforward Sequential Memory Networks,DFSMN)to provide a non-cyclic architecture to model long-term dependencies,By link timing classification(Connectionist Temporal Classification,CTC),as the loss function of the model,builds the end-to-end network architecture of CNN-DFSMN-CTC to construct the acoustic model for Tibetan speech recognition,The 3-gram language model is added to improve the effect of Tibetan speech recognition.The experimental results show that the word error rate of CNN and DFSMN models under the same test set is 31.28% and 28.70%,respectively.The word error rate of the CNN-DFSMN-CTC model was 4.93% and 2.35% lower than the two baseline models,respectively.3.The BERT pre-training model was used to study Chinese-Tibetan machine translation.First,the BERT model(full name: Bidirectional Encoder Representation from Transformers)is trained,and the trained output result is taken as the Input embedding of the Transformer model.Then,the translation was performed through the Transformer model.Finally,the hyperparameters are introduced to optimize the machine translation.Automatic translation quality evaluation method(Bilingual Evaluation Understudey,BLEU)was used to evaluate the translation effect.The experimental results show that under the same test set,the BLEU score after adding the BERT model is 1.3 percentage points higher than the traditional transformer model.The model after hyperparameter optimization has significantly improved the BLEU score compared with the model before optimization.The model translation effect constructed in this article is better than the traditional Transformer model. |
