Research On Key Problems Of Russian-Chinese Military Speech-to-speech Translation Based On Sequence-to-Sequence

Speech translation technology is an important booster for promoting human social exchanges and promoting the progress of human civilization.With the solid advancement of theories such as speech processing and machine translation,and the further development of computer science,the computing power and storage capacity have been further improved.The English-French and Chinese-English and other widely used speech translation systems have been commercialized.However,in the development of speech translation system,the limitations of corpus resources and the lack of bilingual language research,the experiment and application of speech translation in some languages are still in their infancy.In addition,the existing research mostly adopts the cascading speech translation system,and the methods of breaking the key problems are less,and the direct speech translation method without using the intermediate text representation is less used,which brings the research content cumbersome.The translation model is complex and the translation delay is long.Although Russian and Chinese are the first and sixth languages in the world,the application of speech translation has broad prospects.However,Russian-Chinese translation has a late start,and the construction of bilingual-aligned speech corpus resources is not perfect.It is still in the experimental stage.In this research background,this paper is based on the Russian and Chinese military speech translation needs.Based on the analysis and comparison of existing speech translation techniques,this paper designs and implements the end-to-end speech translation model construction method based on the characteristics and differences of Russian and Chinese phonetic pronunciation.Russian-Chinese military language translation system.The main work of this paper and the research results obtained are as follows:(1)According to the characteristics and differences of Russian and Chinese phonetic pronunciation,combined with the spectral map and the Mel cepstral frequency coefficient,the advantages and disadvantages of speech feature representation are adopted.The Mel spectrum is used as the input and output of the speech translation system.Compared with the spectrogram,the Mel spectrum can express the speech frequency characteristics more clearly in the foundation of occupying less storage space.(2)After the Mel spectrum is output,the inverse gray filter is used,and the modified Griffin-Lim algorithm is used to fit the phase of the speech signal.Finally,the signal-to-noise ratio is improved by de-emphasis,and the conversion from the Mel spectrum to the speech waveform is completed.Experiments have shown that the improved vocoder design has a better convergence speed.(3)Combining the characteristics of Russian Chinese pronunciation and the characteristics of two-way long and short memory networks,the model architecture of LSTM-BLSTM is proposed.The Russian Mel spectrum is taken as an input,encoded into a fixed dimension vector,decoded into a Mel spectrum output,and the speech translation process is completed.And optimize the model translation effect by improving the attention mechanism and loss function selection.(4)We collected 49 hours of Russian-Chinese bilingual vocabulary in the military field,designed and constructed a Russian-Chinese bilingual parallel corpus.Based on the Sequence-to-Sequence model,the Russian-Chinese military speech translation prototype system is designed and implemented,and an interactive interface is designed.In the foundation of the design of the corpus and prototype system,the experiment on the translation effect of Russian Chinese sound is carried out.The experimental results show that although the translation accuracy of this system is still insufficient compared with the traditional cascaded speech translation system,the model has the advantages of faster translation speed,good translation speech quality and better translation of short sentences.