Transtron Inc.

Yuki Satomi

Even when a speech is small, the speech is allowed to be detected and an echo is allowed to be appropriately suppressed. Whenever a sample point of a reception signal transmitted through a receiving signal path that transmits a signal to a speaker is acquired, an optimal mask is sequentially generated or selected from base masks as one or a plurality of masks generated based on a learning signal based on a reception signal acquired within a predetermined period before a time point at which the sample point was acquired. Whenever the optimal mask is selected, whether a double-talk state is present is sequentially detected based on a result of comparing an input signal with the optimal mask. When detecting that a speech is not input to a microphone and the reception signal includes a speech, a process of suppressing an echo is sequentially performed on the input signal.

Echo suppression device, echo suppression method, and echo suppression program

There is provided an echo suppression device that allows adding a natural comfort noise even in an environment in which a change in ambient noise is large. An echo is removed from an input signal to generate an echo-removed signal, a noise signal included in the echo-removed signal is estimated to generate an estimated noise signal, and a noise signal included in the echo-removed signal is suppressed based on the estimated noise signal to generate a suppressed signal. A comfort noise is generated based on the estimated noise signal, and the comfort noise is superimposed on the suppressed signal.

A non-linear echo is allowed to be suppressed and voice deterioration is allowed to be suppressed. An echo suppression device for suppressing an echo generated by inputting a voice signal output from a speaker to a microphone includes a compressor on a receiving signal path for transmitting a signal of a receiving signal to the speaker. When a double-talk state has been detected, the compressor performs a compression process on a signal greater than a first threshold among the receiving signals.

An echo is allowed to be effectively removed while a processing load on an arithmetic device is reduced. When a magnitude of a reference signal transmitted to a speaker is equal to or greater than a first threshold, a learning algorithm using nonlinear processing is selected. When the magnitude of the reference signal is smaller than the first threshold, a learning algorithm using linear processing is selected. The echo is learnt using the selected learning algorithm to remove an echo included in an input signal picked up by a microphone based on a learning result.

A double-talk state can be accurately detected, and based on a detection result, echo can be appropriately suppressed. When a sound is output from a speaker and only the output sound is input to a microphone, a comparison is made, for each of different frequency bands, between a frequency mask generated based on a power spectrum or an amplitude spectrum for a learning signal transmitted through a transmitting signal path and a value of a power spectrum or an amplitude spectrum for an input signal input from the microphone, to detect whether there is a double-talk state. In a case of detecting that no signal is being transmitted through the transmitting signal path and that a signal is being transmitted through the receiving signal path, an echo suppressor is used to execute processing of suppressing an echo in the input signal.

Growing community of inventors

Kanagawa, Japan

Yuki Satomi