当前位置：网站首页>Voice conversion history research record

Voice conversion history research record

2022-07-18 22:23:00 【Wsyoneself】

  1. Channel spectrum conversion ： Based on the statistical analysis of the source and target speaker speech , Through parameter mapping
    1. Conversion method based on codebook mapping ： Reduce the number of features of source and target speech through vector quantization , Then convert the centroid vector closest to the source code book into the corresponding target codebook through clustering method to realize speech conversion . shortcoming ： When quantifying, it will cause discontinuity of feature space , And the inter frame information is ignored , The conversion effect is not ideal . Subsequently, a series of improved codebook mapping methods are proposed to solve the discontinuity problem , But it also leads to the problem of over smoothing .
    2. Transformation method based on Gaussian mixture model ： The probability distribution of observation data is expressed by the weighted sum result of a set of Gaussian functions . shortcoming ： Estimate only on the source eigenvector , Instead of joint eigenvector estimation , That is, the inter frame information is not considered enough , Prone to over fitting and over smoothing problems . Subsequently, a large number of mathematical methods have been combined into GMM in , But because of GMM There is a non one-to-one mapping in itself , The resulting over smoothing problem has not been fundamentally solved . Use one GMM To fit the joint distribution of input features and output features , According to the input characteristics and GMM To infer output characteristics .
    3. Transformation method based on Hidden Markov model ： The dynamic change of speech signal can be modeled by using its own implicit state and state transition probability matrix . shortcoming ： because HMM Limited number of implied states for , The dynamic range of voice signal is limited , Thus, the accuracy of conversion processing is restricted
    4. Based on frequency bending conversion method ： Means by stretching or compressing the spectrum along the frequency axis , To adjust the position and bandwidth of the formant , And adjust the energy in each frequency by amplitude scaling , So as to realize the spectrum mapping from the source to the target speaker . characteristic ： It can maintain the naturalness of speech to the greatest extent , And the converted voice quality is high , But the similarity is slightly insufficient , It needs to be combined with other methods to achieve further improvement .
    5. Conversion method based on Neural Network ： Full convolution neural network , Generative antagonistic network , Bidirectional long-term and short-term memory networks are used to realize high-precision conversion from spectral sequence to sequence . More neural networks combine different speech features and adopt different network conversion models . shortcoming ： Current excellent deep learning model , Too many dependent parameters , In the non cooperative mode, when the training data is insufficient , There will be fitting phenomenon , Leading to a sharp decline in performance
    6. Conversion method based on waveform generation ： Directly generate audio waveform sample points , Typical example Wavenet. A deep autoregressive model based mainly on conditional probability modeling , Take various features of speech as conditions , Find a suitable autoregressive model through training . characteristic ： The resulting speech has high clarity and naturalness 、 Good quality and no smoothing problems , But the network generation speed is slow . Aiming at the phenomenon that speech waveform generated by sample points in the network is easy to cause speech collapse , And how to further improve the naturalness of converted speech remains to be further studied .
  2. Prosody Conversion ： Prosody conversion mainly includes the conversion of pitch period 、 Time conversion and energy conversion , The conversion of vocal tract spectrum is expressed as formant frequency 、 Formant bandwidth 、 Spectrum tilt and other conversion .
原网站

版权声明
本文为[Wsyoneself]所创，转载请带上原文链接，感谢
https://yzsam.com/2022/199/202207162002279958.html

边栏推荐

猜你喜欢

随机推荐