#Japan #Tokyo #Asahi Shimbun #Interspeech 2026 #QC-GAN

Asahi Shimbun's Remarkable Achievement in AI Research

Asahi Shimbun Co., led by President Katsuhisa Kakuta, is thrilled to announce that a research paper from its Media Research and Development Center has been accepted at Interspeech 2026, one of the largest international conferences focusing on speech recognition and dialogue technologies worldwide. This remarkable research, spearheaded by lead author Shogo Yamauchi, proposes a novel method for speech enhancement (noise reduction) that dramatically minimizes AI model size while efficiently extracting human voices from noisy backgrounds.

The paper's acceptance into the newly established Long Paper Track signifies the significant impact and academic contributions of this study. This recognition is a testament to the rigorous and innovative research being conducted at the Media Research and Development Center.

In today's world, technologies that enhance voices from surrounding noise have become essential, especially in smartphones, hearing aids, and online meetings. Recent advances in deep learning AI technology have substantially improved the quality of these enhancements, yet achieving high quality often requires larger models that demand extensive computational resources. A critical challenge in this arena is preserving the essential phase information of sound, which often diminishes in smaller models.

To address these challenges, the research employs a mathematical method called quaternion to efficiently manage the size and timing of sound signals. This new method, named QC-GAN, combines superior AI technology for sound processing with the MetricGAN learning approach, which enhances auditory perceptibility or sound quality as perceived by humans.

The results are striking: evaluated using standard datasets, this new model achieves high audio quality while being less than half the size of leading models, with approximately 890,000 parameters. Furthermore, when the model is extremely compact (about 35,000 parameters), it surpasses traditional lightweight methodologies in delivering exceptional quality. Effectiveness has also been validated using datasets closer to real-world usage, demonstrating that quaternion-based techniques can maintain phase accuracy and provide high-quality noise reduction even with minimal computational power.

You can read more about the quaternion concept and its applications in neural networks in Yamauchi's blog here.

The research paper compares the proposed quaternion method (on the right) with traditional real-number-based methodologies (from the paper). While conventional techniques require sixteen parameters for neural network computations (W11 to W44), the proposed method functions efficiently with only four parameters (W1 to W4), leading to a 75% reduction in parameter size.

Asahi Shimbun aims to apply this innovative technique in signal processing for audio and image applications, enhancing the quality of reporting in the field. For instance, it could enable rapid, high-quality noise reduction and transcription of interview audio using fewer computational resources. Additionally, the company’s content creation support service, ALOFA, which offers transcription capabilities, may soon benefit from the adoption of this advanced technology.

About the Research Paper

Shogo Yamauchi, Hideaki Tamori, Makoto Sakai, Yosuke Yamano, Tohru Nitta. QC-GAN: A Parameter-Efficient Quaternion Conformer GAN for High-Fidelity Speech Enhancement. In Proceedings of INTERSPEECH 2026, Sydney, Australia, September 2026. Read the full paper

About Interspeech

Interspeech, organized by the International Speech Communication Association, stands as one of the world's leading conferences for research in speech technology, encompassing areas like speech recognition, synthesis, and dialogue. Each year, numerous papers are submitted from around the globe, undergoing rigorous peer reviews before selection. The acceptance of this paper showcases the international recognition of the Media Research and Development Center's achievements in speech research. Interspeech 2026 will be held in Sydney, Australia from September 27 to October 1.

About the Media Research and Development Center

Established in April 2021, the Media Research and Development Center aims to harness cutting-edge media technologies, including AI, to tackle challenges both within the organization and outwardly. With a focus on advanced research and development in natural language processing and image processing, they strive to address contemporary issues in the media landscape.