Vector quantization based on the dynamic threshold of speaker recognition