The performance of the speaker recognition system declines when training and testing audio codecs are mismatched. In this paper, based on analyzing the effect of mismatched audio codecs in the linear prediction cepstr...The performance of the speaker recognition system declines when training and testing audio codecs are mismatched. In this paper, based on analyzing the effect of mismatched audio codecs in the linear prediction cepstrum coefficients, a method of MAP-based audio coding compensation for speaker recognition is proposed. The proposed method firstly sets a standard codec as a reference and trains the speaker models in this codec format, then learns the deviation distributions between the standard codec format and the other ones, next gets the current bias via using a small number adaptive data and the MAP-based adaptive technique, and then adjusts the model parameters by the type of coming audio codec format and its related bias. During the test, the features of the coming speaker are used to match with the adjusted model. The experimental result shows that the accuracy reached 82.4% with just one second adaptive data, which is higher 5.5% than that in the baseline system.展开更多
文摘The performance of the speaker recognition system declines when training and testing audio codecs are mismatched. In this paper, based on analyzing the effect of mismatched audio codecs in the linear prediction cepstrum coefficients, a method of MAP-based audio coding compensation for speaker recognition is proposed. The proposed method firstly sets a standard codec as a reference and trains the speaker models in this codec format, then learns the deviation distributions between the standard codec format and the other ones, next gets the current bias via using a small number adaptive data and the MAP-based adaptive technique, and then adjusts the model parameters by the type of coming audio codec format and its related bias. During the test, the features of the coming speaker are used to match with the adjusted model. The experimental result shows that the accuracy reached 82.4% with just one second adaptive data, which is higher 5.5% than that in the baseline system.
