A voice transmission device has a tandem pass through function in an STM, ATM, IP network, and a noise canceller (21) is provided at a latter stage of an echo canceller (12) in a coding part, and in a case where a multistage connection state does not occur, noise removal is performed to carry out efficient coding, and in a case where a relay is performed in the multistage connection state, switching is performed to stop the operation of the noise canceller (21), and voice deterioration by redundant decoding and coding is prevented, and therefore, even in a cellular phone having no noise cancel function, or the like, and at both a normal time and a time of a tandem pass through state, high quality voice transmission can be performed, and further, since an unvoiced portion in voice data is increased by removing a noise component, a portion subjected to coding becomes small, and transmission amount of the line is reduced.