Evaluation of overlapping vowels with information geometry - Part 1 (Japanese)

The detection of synthetic speech generated by artificial intelligence has become a growing challenge for sectors such as cybersecurity and biometric authentication. While generative systems produce utterances based on limited training spectra, natural human voice exhibits a spectral richness that reflects the flexibility of the vocal tract. An emerging approach, inspired by information geometry, proposes analyzing the distributions of vowels —key phonemes in individual identification— using Wasserstein metrics and persistent homology. In languages such as Japanese, with a syllabary of only five vowels that correspond uniquely to specific sounds, applying this comparison is simpler: by normalizing the spectra as probability density functions over the frequency bands perceived by the cochlea, it is observed that the Wasserstein distances between synthetic vowels are much shorter than natural ones. This property allows, through topological mapping, decomposing the distributions into differentiable clusters. The technique not only has forensic applications but also opens doors to developing more robust identity verification systems. In this context, companies like Q2BSTUDIO offer artificial intelligence solutions for businesses that integrate advanced spectral analysis models, as well as AWS and Azure cloud services to process large volumes of audio data in real time. The combination of custom software with information geometry techniques enables organizations to develop custom applications for detecting voice deepfakes. Additionally, the incorporation of AI agents and Power BI tools facilitates the visualization of spectral clusters and the generation of business intelligence service reports. This approach strengthens cybersecurity by preventing impersonation through synthetic voice, a growing risk in corporate environments. Ongoing research promises to extend the method to other languages and contexts, consolidating information geometry as a key tool in the fight against voice fraud.