LG Demonstrates Competitiveness at 'Interspeech 2022', the Premier Conference in the Speech Processing

“Interspeech 2022,” an international academic conference on AI speech and signal processing, was held at Songdo Convensia in Incheon from September 18 to 22. This year marks the 23rd anniversary of Interspeech, the largest academic conference in the field of speech processing, hosted by the International Speech Communication Association (ISCA). Speech signal processing, speech recognition, speech synthesis, multilingual translation, speech conversion, and assistive speech therapy, all of which are required for human-to-human, human-to-object, and object-to-object language and voice communication, were discussed this year under the theme “Human and Humanizing Spoken Language.”

This year, LG presented voice recognition technology from the LG AI Research, LG Electronics, and LG U+ at an integrated booth. Jungkyu Choi, Language Lab Leader of LG AI Research, spoke at the Industry Talk and presented the voice and language technology of LG's super-giant AI EXAONE.

LG AI Research demonstrated the YouTube news video STT (speech to text) technology at the LG integrated booth. STT is a function that extracts keywords while simultaneously converting the audio of a YouTube news video into text in real-time. The application of this technology to customer service can increase efficiency by expediting the preparation of response manuals and the connection to experts in the relevant field. LG AI Research also introduced a technology that captures the essence of voice consultations. Using this technology, it is possible to categorize the contents of consultations, such as complaints and refunds, as well as the emotional state of the inquiring customer as positive, negative, or neutral.

LG Electronics gave demonstrations of home appliances with voice recognition capabilities. The launch of LG Electronics’ “Puricare Object Collection Water Purifier,” the first water purifier in Korea equipped with a voice recognition function, garnered significant attention last month. Puricare Object Collection Water Purifier is equipped with a fine water extraction function that delivers the exact amount of water requested, allowing you to obtain water without pressing a button by saying “Hi LG!” and “Give me 550mL of cold water.” By saying “Once more,” the device will release the same amount of water again. In addition, a new refrigerator, which is scheduled for release at the end of September, utilizes voice recognition to control panel effects and doors. If you say “Set to Christmas colors,” the refrigerator's appearance changes accordingly, and if you say “Open the door,” the door opens automatically, which is expected to increase user convenience.

For media content search and customized services, LG U+ introduced AI-based speech recognition, speaker recognition, and natural language processing (NLP) technologies. It recognizes voice through the microphone built into the remote control and recommends customized content from the IPTV service using STT technology and NLP technology. Not only that, but it also recommends age-appropriate content by differentiating between infant and adult voices, and restricts access to content that is not age-appropriate.

Jungkyu Choi, Leader of LG AI Research's Language Lab, presented and discussed EXAONE's advancements in speech and language technology during the conference's Industry Talk event. Self-supervised learning, E2E (End-to-End) voice technology, optimization, and domain application are the four phases of EXAONE voice research. Choi stressed the significance of the individual research phases, stating that the research into optimization enables the technology to be implemented with greater precision, faster inference, and lower computing costs. In addition, he revealed that LG AI Research has developed a GPU-less streaming E2E STT inference technology that operates in real-time, as well as positive research outcomes of EXAONE as a result of the internal evaluation, namely that it had achieved stable voice recognition accuracy.

At the conference, LG Electronics presented three papers during the poster session. The first two papers, QbyE-MLPMixer: Query-by-Example Open-Vocabulary Keyword Spotting using MLPMixer and DyConvMixer: Dynamic Convolution Mixer Architecture for Open-Vocabulary Keyword Spotting, discuss “user-defined wake-up-word (WUW) recognition,” which can register and recognize voice recognition WUWs such as “Hi LG!” as desired word combinations. The third paper, titled Online Learning of Open-set Speaker Identification by Active User-registration, focuses on the technology that distinguishes a speaker's voice, which can be used for customer-customized voice recognition services.

LG Group, including LG AI Research, is conducting research on voice processing technology in a variety of field papers at more conferences, operating booths, and interacting actively with the academic community to further strengthen research activities.