HKUST Xunfei Exclusive Founding Sponsored International Speech Conference, 14 papers were included – Programmer Sought

This year is the second time for INTERSPEECH to come to China.

INTERSPEECH is a top-level international conference sponsored by the International Speech Communication Association (ISCA), and it is also recognized as one of the two top conferences in the field of speech in the world.

This year is the 21st INTERSPEECH conference. Thanks to the outstanding contributions and active efforts of Chinese scholars in the field of speech, INTERSPEECH 2020 will come to China for the second time after it was successfully held in Beijing in 2000.

INTERSPEECH 2020 has received enthusiastic attention from scholars around the world. A total of 2,140 valid paper submissions have been accepted, and a total of 1,022 papers have been accepted, covering speech, signal processing, spoken language processing and other aspects.

14 papers of HKUST iFLYTEK were included in the top conference, once again demonstrating the source of technological innovation.

Relay on source technology innovation

Among the papers included in INTERSPEECH 2020, iFlytek Research Institute of HKUST and the Speech Laboratory of the University of Science and Technology of China published a total of 14 papers, including speech recognition, speech synthesis, speech enhancement, speech emotion recognition, sound event detection, speaker recognition Innovation in multiple technical directions.

In the direction of speech recognition, iFLYTEK focuses on speaker adaptive technology research. This is a key technology in speech recognition. The purpose is to allow the algorithm model to quickly adapt to the different pronunciation characteristics of each speaker, so as to achieve better recognition results.

In response to this technical problem, the R&D team has proposed a number of technological innovations, one of which is a method based on the Cascading Attention-over-Attention mechanism, which significantly improves the effect of online speaker adaptation.

In the direction of speech synthesis, non-parallel speech conversion tasks are hot and difficult issues in academic research. The goal of speech conversion is to process the input source speech to make the output sound like the target person’s speech, and keep the semantic information unchanged in the process. It has a wide range of application scenarios in personalized speech synthesis, entertainment applications, and voice anonymization.

The R&D team jointly optimized the recognition synthesis model and introduced adversarial learning objectives to decouple the semantic features and the speaker’s timbre features, thereby improving the similarity of speech conversion.

In the direction of speech enhancement, the R&D team combined deep learning technology with traditional microphone array algorithms, and achieved excellent results in speech separation and recognition in the cocktail party scene of CHiME-6 (CHiME, an international multi-channel speech separation and recognition competition). Significant improvement.

In the direction of speaker recognition, how to obtain accurate speaker labels has always been a difficult point in speaker recognition. The R&D team proposed a method combining speaker confirmation and speaker classification, which reduces the dependence of the algorithm model on accurately labeled data and improves the accuracy of speaker recognition.

These original technological innovations will further enhance iFLYTEK’s intelligent voice capabilities, not only empowering products such as iFLYTEK translators, iFLYTEK hearing, and XFLYTEK learning machines, but also make it continuously optimized and iterative, bringing more convenience to users , better services; it can also empower industries such as medical care, finance, justice, and education, and promote innovative applications in the A.I.+ industry.

Supporting the successful holding of two major voice competitions

As one of the satellite conferences of INTERSPEECH 2020, the international speech synthesis competition Blizzard Challenge 2020

A joint workshop with Voice Conversion Challenge 2020, an international voice conversion competition, will be held on October 30th. HKUST iFLYTEK and USTC National Engineering Laboratory for Speech and Language supported and participated in these two competitions.

The Blizzard Challenge is an international speech synthesis competition jointly initiated by research institutions in the United States, Japan, and the United Kingdom. It is the largest and most influential top competition in the field of speech synthesis in the world today. HKUST Xunfei has continuously achieved technological breakthroughs in previous competitions and won the championship many times. This year, iFLYTEK provided training corpora of Mandarin Chinese and Shanghainese for the Blizzard Challenge, leading the technological development of the industry.

Voice Conversion Challenge is also the most professional and top speaker voice conversion competition in the world today. In this year’s competition, the system jointly submitted by iFLYTEK and the National Engineering Laboratory of Speech and Language Information Processing of the University of Science and Technology of China won the first place in 8 indicators and 7 items in the two tasks of “same language conversion” and “cross language conversion”. One, successfully defended the title.

I believe that as the two major competitions continue to be held, they will continue to attract the attention of cutting-edge researchers in the field of artificial intelligence and lead the development of intelligent voice technology.

Exclusive Founding Sponsor Summit

The 21st INTERSPEECH2020 International Conference will be held in Shanghai. Affected by the new crown epidemic, the summit meeting originally planned for October 25-30, 2020, will be converted to an online meeting held at the same time. HKUST Xunfei became the exclusive Founding sponsor of this summit, supporting the smooth landing of the summit in China.

During the conference, iFLYTEK will hold an online “Enterprise Forum” from 19:15 to 20:15 on October 28. Dr. Gao Jianqing, vice president of A.I. Today and Tomorrow” theme report, sharing iFLYTEK’s experience in voice technology research and development and product innovationThe latest developments in new areas and application cases of AI-enabled industries, and an outlook on the next development trend of voice technology.

Source of this article: Popular News

Responsible editor: Chen Tiqiang_NB6485