Changeling Lab introduced the PhoneticXeus model for speech recognition

Changeling Lab team released PhoneticXeus — a specialized model for converting audio streams into International Phonetic Alphabet (IPA) phonemes. The solution is focused on linguistic analysis and automatic transcription tasks with support for over 70 languages.

The model's architecture is optimized for high accuracy of sound unit recognition in multilingual conditions. During primary testing, the system demonstrated a result of 13 points when processing data for 3 hours of work. The developers provided access to the model's weights and code for reproducing the results through the Hugging Face platform.

An interactive browser demo version is available for familiarization with the system's capabilities. The tool allows users to upload their own audio files to receive phonetic markup in real-time, simplifying the integration of the model into research projects on computer linguistics and speech recognition.

The release of PhoneticXeus expands the toolkit for working with phonetic data, which traditionally requires significant computational resources. The use of IPA standardization allows for the unification of audio data processing obtained from various language environments, ensuring the comparability of results in cross-language analysis.