Abstract
Researchers at NTT have developed a suite of media processing technologies designed to bridge the gap between noisy real-world environments and digital productivity tools. By combining an "intelligent microphone" that utilizes hybrid beamforming with a high-speed Deep Neural Network (DNN) speech recognition engine, they have enabled accurate, real-time transcription and meeting support even in extreme 100-dB noise environments.
Introduction
While Information and Communication Technology (ICT) has streamlined many business tasks, two major "input barriers" persist. First, environmental noise in locations like factories, construction sites, or moving vehicles often renders traditional speech recognition ineffective. This forces workers to rely on manual keyboards, which is frequently impossible when hands are occupied or a workspace is unavailable. Second, the "meeting tax"—the time-consuming and error-prone process of manually recording minutes—often prevents participants from fully engaging in discussions.
The proposed system aims to overcome these barriers by segregating target human voices from chaotic backgrounds and processing them into text instantly, eliminating the lag typically associated with complex neural networks.
Technical Approach
The system's architecture relies on three core technical components:
- The Intelligent Microphone: This system employs a hybrid approach, combining spatial beamforming with spectral filtering. By analyzing spatial power distribution, frequency, and temporal fluctuations, the system estimates a spectral filter that suppresses noise while preserving the target voice.
- Advanced Voice Activity Detection (VAD): To prevent premature cut-offs or post-utterance lag, a VAD technique identifies the "temporal density" of speech. It uses a specific speech activity model to distinguish between brief pauses within a sentence and the actual end of an utterance.
- VoiceRex NX2014 Engine: A real-time speech recognition engine powered by Deep Neural Networks (DNNs). Engineering optimizations ensure high-speed processing while maintaining the superior accuracy of deep learning.
The system was developed and tested using real-world industrial data from factories and highways, as well as governmental records from the National Diet of Japan and local assembly meetings.
Results and Discussion
The system demonstrates remarkable noise robustness, achieving highly accurate voice recognition in environments with noise levels up to 100 dB. A key technical achievement is the ability to reduce ambient noise power by a factor of 1/10,000 with minimal degradation to the target voice. Furthermore, the new VAD technology reduced detection errors by at least 15% compared to conventional methods.
Beyond simple transcription, the researchers view speech recognition as a tool for "conversation promotion." The RexMeet application utilizes these technologies to provide real-time support during meetings.
By generating "tag clouds" of high-frequency words and allowing real-time tagging of important comments, the system helps participants organize thoughts during the discussion. It also provides behavioral feedback, such as monitoring speaking speed and participation levels, which helps participants improve their communication habits and overall meeting productivity.
Conclusion and Future Work
The current research successfully addresses the audio and speech recognition challenges of business support in noisy environments. However, for truly intelligent support, the system must evolve to understand context beyond words. Future work will focus on media integration—combining speech recognition with natural language processing (NLP) and image recognition—and developing contextual awareness to provide proactive assistance based on the user's environment and focus.
Summary updated