Voice conversion mainly involves technical records

Speech analysis and synthesis , At present, the main analytical and synthetic methods ：

1. Harmonic plus noise model (HNM)： Divide the signal into harmonic components and noise components , The harmonic component represents the low-frequency part of the signal , Can be determined by the fundamental frequency 、 Range 、 Phase is represented by three parameters ; Noise represents the high-frequency part of the signal , It can be represented by Gaussian white noise through high pass filter , Then use the specific algorithm to calculate the amplitude and phase values .
2. STRAIGHT( Speech transformation and representation using weighted spectral adaptive interpolation ): A pitch adaptive time-frequency spectrum smoothing algorithm model is proposed , It can reduce the interference between signal period and spectrum
3. Speech feature parameter extraction , At present, the mapping parameter features extracted in speech conversion are mainly the local features of segmental information and the context features of suprasegmental information . The local feature is mainly spectral envelope 、 Cepstrum, formant and other parameters , More commonly used are line spectrum pairs (LSF) Parameter and Mel frequency cepstrum coefficient considering human auditory characteristics (MFCC), Context feature mainly refers to the dynamic information between speech frames .
4. Voice time alignment ：
   1. For parallel corpus ： The most commonly used method is dynamic time warping (DTW), Calculate the best time alignment for each utterance pair or between each phoneme pair . After dynamic time domain warping , The final result is a pair of source and target feature sequences of equal length .
   2. For non parallel corpus ： Can be used based on WaveNet Voice conversion method of silent coder , This method does not need to deal with intermediate features , But the use of wavenet Directly map the speech posterior to the waveform samples , This avoids the estimation error caused by vocoder and feature conversion .
5. Evaluation of conversion effect ：
   1. objective evaluation ： Based on the measurement of speech data distortion , Some distance criterion is used to measure the similarity between the converted speech and the original target speech , From this, we can get the evaluation method of the advantages and disadvantages of the conversion method . The main objective evaluation indicators have mean square error (MSE)、 Spectrum loss really ( SD) and mei Er pour Spectrum loss really (MCD),MSE、SD and MCD The smaller the value of , The smaller the distortion , The higher the conversion accuracy .
   2. Subjective evaluation ： Taking people as the main body , Test the voice through people's subjective feelings . Compared with objective evaluation , Subjective evaluation results are more reliable . Subjective methods are generally based on speech quality and speaker feature similarity 2 From two angles , The method adopted is mainly average opinion score (MOS) and ABX:
      1. MOS test ： Let the evaluator base on 5 Score the subjective feeling of the test speech by three grades , It can be used for subjective evaluation of speech quality , It can also be used to evaluate the similarity of speaker features .MOS Score is the comprehensive average result of all test statements and all evaluators .
      2. ABX test ： It mainly evaluates the conversion effect based on the speaker feature similarity of the converted speech , It draws lessons from the principle of speaker recognition . During the test , The assessors listen separately 3 Paragraph voice A、B and X, And judge the voice in terms of personality characteristics A still B Closer to X.(X Is the voice after conversion , A and B They are source speech and target speech . Finally, the judgment results of all evaluators are counted , Calculate the percentage that sounds like the target voice .