Data Analytics Lab Unveils Research on AI-Generated Voice Detection Technology

Data Analytics Lab's Research on AI Voice Detection Technology

Data Analytics Lab, based in Chuo, Tokyo, has announced its latest findings on voice detection technology for AI-generated content. This research was conducted in collaboration with Evixer as part of a project approved by the Japanese Ministry of Internal Affairs and Communications aimed at tackling the issues of misinformation on the internet.

Overview of the Research

This project merges Evixer's advanced sound signal processing with Data Analytics Lab's AI and data analysis expertise. It focuses on enhancing countermeasures against the rising threat of misinformation, particularly deepfakes and fake audio content.

Key Research Outcomes

The research produced a framework for analyzing AI-generated audio content. This includes:

1. Building a Wide-Verse Verification Platform
To analyze synthetic voices, the research examined various advanced voice generation models, including:
- Tortoise
- XTTS (Multilingual Model)
- Qwen3-TTS
This comprehensive examination enables the evaluation of multiple generation techniques.

Notably, with models like XTTS, which are based on large-scale multilingual learning, the testing was conducted under conditions that closely mimic real-world AI environments.

2. Systematic Generation and Feature Analysis of Synthetic Voices
The project entailed:
- Organizing and systematizing the conditions under which synthetic voice data is generated
- Analyzing audio signals (such as spectrograms)
- Identifying structural differences between synthetic and natural voices
This facilitated a quantitative grasp of the characteristics of AI-generated voices, laying the groundwork for the development of universally applicable detection technologies.

3. Evaluation of Synthetic Voice Detection Models Using Deep Learning
The team performed:
- Investigations and tests of deep learning models
- Construction of training datasets for voice detection
- Development of processes to evaluate detection accuracy
As a result, the effectiveness of the detection model, which leverages specific features of AI-generated voices, has been validated to a certain degree.

4. Fusion of Acoustic Signal Processing and AI
By integrating Evixer's sound signal processing expertise with AI techniques, the project contributed to strengthening Evixer's synthetic voice detection system (EAF). Key activities included:
- Generating synthetic voice data and conducting operational tests of diverse generation models
- Analyzing features of audio signals to understand differences between synthetic and human voices
- Validating detection accuracies and building evaluation datasets using deep learning models
This collaborative effort enhanced the technical insights necessary for improving detection accuracy in EAF's systems.

Positioning of this Research

The project, led by Evixer and supported by the Ministry of Internal Affairs, sees Evixer responsible for developing acoustic signal processing and AI countermeasure technologies, while Data Analytics Lab focuses on data design, analysis, and verification.

Data Analytics Lab specifically handled:

- Generation and design of synthetic voice data
- Analysis and extraction of audio signal features
- Validation of detection models and establishment of evaluation frameworks

This work significantly contributes to the overall technical enhancement of the research initiative.

Societal Significance

As the advancement of generative AI presents challenges to the trustworthiness of audio and visual content, this research addresses a significant societal issue. Its findings are positioned to assist in:

- Countering deepfake technologies
- Verifying information authenticity
- Supporting AI governance and standardization efforts

The insights gained from this research could advance technological development in these critical areas.

Future Developments

Looking ahead, Data Analytics Lab aims to continue collaborating with Evixer, leveraging both companies' strengths to further the advancement of acoustic signal processing and AI solutions to combat misinformation effectively.

About Data Analytics Lab

- Website: dalab.jp
- CEO: Masahiko Kondo
- Address: 5-4-18 Tsukiji, Chuo, Tokyo, Shiodome East Side Building 6F
- Established: April 26, 2019
- Capital: 20 million yen
- Services: Data analysis support, AI development, and data talent education services