Currently, the system can only successfully register one voice, and that voice must be the loudest one at registration, but the team's goal is that the system will still work even if the loudest voice in a particular direction is not the target person.
Sefik Emre Eskimez, a senior researcher at Microsoft who studies speech and artificial intelligence, said it is very difficult to capture a voice in a noisy environment. He was not involved in the study.
"I know a lot of companies want to do that." "If they can do that, it will unlock a lot of applications, especially for meetings," he said.
Samuele Cornell, a researcher at Carnegie Mellon University's Institute for Speech Technology in the US, believes that while speech separation research tends to be theoretical rather than practical, the work has clear applications in the real world.
He was not involved in the study. But he said: "I think it's a step in the right direction and a very novel experiment."Scientists train diverse data sets based on deep learning to achieve non-orthogonal multiplexing of multimode fibers
This study provides a new way to overcome the mode dispersion of multi-mode fiber and give full play to the advantages of high-throughput transmission of multi-mode fiber.
It is expected that scholars in the field will pay more attention to the high-throughput transmission medium of multi-mode fiber, and attract more interdisciplinary scholars to use deep learning methods to explore the application of non-orthogonal multi-dimensional multiplexing transmission in different fields.
In addition, the research also provides a new perspective for the field, that is, with the support of data-driven AI algorithms, to arouse the attention of optics and even the entire field of information science to deep learning.
Qin Yuwen believes that many things that previously seemed impossible or needed more harsh conditions to achieve are expected to break through under the ability of AI.
Pan Tuqiang and Ye Jianwei, master students at Guangdong University of Technology, are co-first authors, and Professors Qin Yuwen and Xu Yi are co-corresponding authors.
At this stage, optical communication focuses on the direction of large-capacity optical fiber communication, and breaks through the transmission capacity of single fiber through advanced modulation algorithms and AI.
At present, division multiplexing is one of the most important multiplexing modes in single-mode optical fiber communication, including wavelength division multiplexing, partial division multiplexing and space division multiplexing.
This technology depends on the physical orthogonality between channels, and if the orthogonality between channels deteriorates, it will greatly increase the complexity of digital signal processing in the receiver.
In the multiplex transmission technology of single-mode fiber, the basis of multichannel multiplexing is that there is physical orthogonality between each multiplexed channel.
Based on this, the researchers asked, is it possible to replace a transmission medium with a larger transmission flux?
Multi-mode fiber is very common in daily life as an information transmission medium, for example, short-range optical interconnection inside large data centers can use multi-mode fiber.
In previous studies, the team has applied deep learning to single-mode fiber communication to improve the performance of fiber transmission systems, but deep learning still cannot solve the non-orthogonal multiplexing problem in single-mode fiber communication.
The researchers said: "Because our research goal is to increase the transmission capacity of a single fiber, so support thousands of transmission modes, multi-mode fiber with high throughput transmission characteristics become the research object of breaking the orthogonal multiplexing paradigm."
In a preliminary attempt, they changed the polarization of the input signals so that the polarization of the two input signals was not perpendicular to each other.
Experiments show that the deep neural network can decode two non-orthogonal signals in this case.
Further, the team made the wavelength, spatial position and polarization state of the two channels exactly the same.
In this case, the method of deep learning can still achieve high-fidelity demultiplexing of two signals, which provides the basis for high-flux non-orthogonal multiplexing transmission based on multi-mode fiber.
Notably, the research relies not only on deep learning's powerful end-to-end mapping capabilities, but more importantly, the physical properties of the multimode fiber itself.
