In scenarios such as remote work, open offices and call centers, multiple people may simultaneously have independent spoken interactions with their devices in the same room. The speech of competing speakers will however be picked up by all microphones, both reducing the quality of audio and exposing speakers to breaches in privacy. We propose a cooperative cross-talk cancellation solution breaking the single active speaker assumption employed by most telecommunication systems. The proposed method applies source separation on the microphone signals of independent devices, to extract the dominant speaker in each device. It is realized using a localization estimator based on a deep neural network, followed by a time-frequency mask to separate the target speech from the interfering one at each time-frequency unit referring to its orientation. By experimental evaluation, we confirm that the proposed method effectively reduces crosstalk and exceeds the baseline expectation maximization method by 10 dB in terms of interference rejection. This performance makes the proposed method a viable solution for cross-talk cancellation in near-field conditions, thus protecting the privacy of external speakers in the same acoustic space.

The use of speech source localization (SSL) and its applications offer great possibilities for the design of speaker local positioning systems with wireless acoustic sensor networks (WASNs). Recent works have shown that data-driven front-ends can outperform traditional algorithms for SSL when trained to work in specific domains, depending on factors like reverberation and noise levels. However, such localization models consider localization directly from raw sensor observations, without consideration for transmission losses in WASNs. In contrast, when sensors reside in separate real-life devices, we need to quantize, encode and transmit sensor data, decreasing the performance of localization, especially when the transmission bitrate is low. In this work, we investigate the effect of low bitrate transmission on a Direction of Arrival (DoA) estimator. We analyze a deep neural network (DNN) based framework performance as a function of the audio encoding bitrate for compressed signals by employing recent communication codecs including PyAWNeS, Opus, EVS, and Lyra. Experimental results show that training the DNN on input encoded with the PyAWNeS codec at 16.4 kB/s can improve the accuracy significantly, and up to 50% of accuracy degradation at a low bitrate for almost all codecs can be recovered. Our results further show that for the best accuracy of the trained model when one of the two channels can be encoded with a bitrate higher than 32 kB/s, it is optimal to have the raw data for the second channel. However, for a lower bitrate, it is preferable to similarly encode the two channels. More importantly, for practical applications, a more generalized model trained with a randomly selected codec for each channel, shows a large accuracy gain when at least one of the two channels is encoded with PyAWNeS.

No abstract