Industry Talk
home
Industry Talk
- Title: Is there money in speech generation research?
Monday 19 September 13:30 - 14:10 KST at R107-109Abstract You have reached SOTA in a speech generation task, whether that be text-to-speech, voice conversion, voice cloning, etc. Now what? Ever wonder how your research would translate to real world impact, especially in a for-profit setting? In this talk, we explore current market opportunities in speech generation research for current SOTA research, as well as some future market opportunities for when technology advances further.
Charlie Choi
LOVO,
Founder - Title: The All-Neural Era: Progress in Speech Recognition Technology at Alexa
Monday 19 September 13:30 - 13:50 KST at R104-106Abstract Speech recognition technology is completing a dramatic shift, the move to an all-neural architecture replacing the conventional stack of disjointly-trained neural and non-neural sub-systems. The unified all-neural architecture improves the SOTA accuracy, while also achieving superior memory and compute compression, enabling streaming low-latency speech recognition at the edge, where resources are constrained. The unified all-neural architecture, in conjunction with multi-task learning approaches, overhauls the boundary between the spoken language understand (SLU) sub-components of voice AI assistants, including endpointing, speaker segmentation, device-directed classification, language identification and language understanding, allowing for a jointly-trained unified stack that goes beyond a transcription-only task. In this talk, we give an overview of challenges and our recent progress in addressing different aspects of adapting all-neural technology for production ASR. Namely, we highlight techniques such as semi-/weakly-supervised learning, incremental learning (to learn new trends and usage distribution shift), contextual biasing, and multi-task learning to bring the benefits of all-neural architecture to Alexa, the AI assistant that serves millions of customers daily.
Bjorn Hoffmeister
Amazon
Director Applied Science, Alexa Speech
Ariya Rastrow
Amazon
Sr. Principal Scientist, Alexa Speech - Title: A Unified Voice Synthesis Framework with High Controllability and Modularity for Media Production
Tuesday 20 September 12:30-12:50 KST at R104-106Abstract Recent advancements in generative models are changing the process of media production. Creators can now simply type in a few words to get a clear image of their ideas or use automated services for previously labor-intensive work like color grading and audio mastering, cutting months of work down to a few minutes. In speech synthesis, end-to-end approaches have reached near-human quality by addressing the training-inference mismatch problem of the previous two-stage framework. For generative models to be actively adopted in industry-grade media productions, such as music, movies, and games, however, it is essential to provide creators with models that can more accurately reflect their artistic intent. In this talk, we will share our design principles for constructing a unified voice synthesis framework. By following the proposed framework, we expect creators to find the controllability for pinpointing specific styles of expression, as well as the modularity to handle multiple layers of input. Lastly, short demos of our approach will be shown to introduce how the proposed analysis and synthesis framework can be extended to various applications in the media industry.
Hoon Heo
Supertone
Co-founder & CTO
Hyeong-Seok Choi
Supertone
Co-founder & Research Lead - Title: Out-of-Domain Decision Algorithm for Result Processing of Multiple Speech Recognition Systems Tuesday 20 September 12:30 - 12:50 KST at R107-109Abstract Automotive OEMs provide speech recognition functions by collaborating with external ICT companies. The speech recognition functions provided by these ICT companies perform well in daily activities such as playing musics, checking weathers and etc. But it does not perform well in vehicle-specialized circumstances. We developed vehicle-specific speech recognition system but there was a limitation on the supply of contents embeded to the system. To mitigate the limitation, we could combine the vehicle specific speech recognition system with the one that is provided from ICT companies. However, there is a difficulty on deciding which system to be used for the final result. We would like to present the ways of combining the two different speech recognition system seamlessly in this session.
Sungsoo Park
Hyundai Motor Company
Senior Research Engineer, Voice Recognition Development Team - Title: Tackle How to Identify and Understand High Qualitative AI Data Solutions Will Improve Your Model Performance Tuesday 20 September 12:50-13:10 KST at R104-106Abstract Intelligent speech interaction technology starts with ASR technology and aims to achieve human-machine dialogue. It includes ASR, NLU & NLG technology and TTS technology. Datatang owns 200,000 hours off-the-shelf speech datasets, covering multiple languages, equipments, environments, and types, and supports customized data service to help customers quickly improve ASR models.
Pan Pan
Datatang
Head of Overseas Department - Title: From ASR Research Idea to Production: Speech Recognition in the Real World Tuesday 20 September 12:50-13:10 KST at R107-109Abstract We are in the middle of an explosion of novel ideas and innovations in the field of deep learning and speech recognition. However, taking many of these innovation ideas into the field is not trivial. ASR systems in production are faced with many challenges we often do not discuss in academic environments such as latency, memory efficiency, quality, privacy, rejection, versioning, personalization, etc to name a few.
In this talk I'll try to give an overview of these challenges educating students about the hard realities of ASR systems in real production environments.
Pedro Moreno Mengibar
Google
Senior Research Director - Title: EXAONE(Expert AI for Everyone) in Speech & Language fields Wednesday 21 September 12:30 - 12:50 KST at R104-106Abstract In this session, we will introduce LG AI Research's super-giant AI, EXAONE. LG AI Research has been developing EXAONE (Expert AI for Everyone) that AI can help experts more professionally. We are developing EXAONE that can achieve the best results with zero shot and few shot learning by overcoming scalability issues, which are the limitations of existing AI technology, and minimizing the fine tuning process that has required a lot of effort. In addition, we will introduce the current status and direction of EXAONE voice research, such as the methodology of securing various real voice data, understanding the context of speech, and optimization for practical uses.
Stanley Jungkyu Choi
LG AI Research
Lab Leader / Vice President - Title: AI Research at Kakao Enterprise Wednesday 21 September 12:30 - 12:50 KST at R107-109Abstract Kakao Enterprise is a Korean technology company that does researches on search, cloud computing, big data analysis, and artificial intelligence. Based on the above technologies, we provide various services such as messenger-based business work platform, chat bot, contact center solution and so on. Using all the power of AI technology, Kakao Enterprise aim at connecting all elements and solving inconveniences in everyday lives. We also like to create new lifestyles by focusing on better values. In this presentation, we mainly focus on AI research and services at the company. Our AI technologies deal with whole kinds of multi-media data like speech, image, video, and text, and research areas we do are speech recognition, speech synthesis, computer vision, natural language processing, translation, and OCR. In this short talk, we briefly introduce our AI technologies and applications, but since the main focus of Interspeech is speech-related research, we take a look at speech recognition and TTS in more detail.
Chul Min Lee
Kakao Enterprise
Leader of STT Engine Development - Title: Towards Interactive AI Wednesday 21 September 12:50 - 13:10 KST at R104-106Abstract KT is leading Korean IPTV services linked to “GiGA genie”, KT’s speech recognition AI based smart speaker. In addition, KT’s interactive AI has enabled AI contact center, which automates customer service center in various industries including finance.
Given KT’s base model provided, partners need to upgrade the model with additional data. Therefore, data-centric approach is a key essential component addressing the challenge. In this talk, I will present how KT empowers partners with Large AI model, MI:DEUM and interactive Language Experiment Tool Suite (LETS).
Soonmin Bae
KT
Senior Vice President, AI2XL, Institute of Convergence Technology - Title: CLOVA Speech Technology and Services Wednesday 21 September 12:50 - 13:10 KST at R107-109Abstract Abstract: In this industry talk, I will introduce the techniques CLOVA has and how these techniques are applied to actual service by going over CLOVA’s main services.
Also, I would like to talk about our speech recognition model. How far our model has come, what kind of constraints we are trying to solve, and how we are trying to proceed our work.
Chankyu Lee
NAVER
Leader / NAVER CLOVA SPEECH - Title: Latest Voice Technologies and their Applications at Samsung Research Thursday 22 September 12:30 - 12:50 KST at R104-106Abstract Abstract: In this industry talk, we give an overview of various voice-related research activities at the Language and Voice Team (LVT) at Samsung Research, which is the central research institute of the Device eXperience (DX) division of Samsung Electronics. In the recent years, we have been conducting cutting-edge research on various voice technologies for a wide range of Samsung products including mobile phones, digital televisions, wearable devices, home appliance devices, and so on. Especially, we have been focusing on on-device fully neural Automatic Speech Recognition (ASR) and Text-To-Speech (TTS) technologies, which show performance comparable to that of server-side solutions using the latest end-to-end technologies. In this talk, we review the architecture of our ASR/TTS models, compression techniques, and their applications in our products. We also briefly discuss our future research direction.
Changwoo Han
Samsung Research
Principal Engineer
Jonghoon Jeong
Samsung Research
Principal Engineer - Title: Speech Synthesis Technology Disrupts Ways of Content Creation Thursday 22 September 12:50 - 13:10 KST at R104-106Abstract Typecast is a service that empowers content creators to create videos and realistic voice-overs with AI actors. By applying cutting-edge technology in speech synthesis, it innovates the way content is created. To deliver AI actors, we focus on two problems. First, achieving results comparable to the quality of human actors. Second, using AI to enable previously impossible tasks. In this talk, we will share how we have integrated the latest speech synthesis technology into the product, which already has more than a million users. We will also talk about what kind of research is needed in this industry in the future.
Younggun Lee
Neosapience
AI Research Lead
Sponsor
Supported
by
by
