The goal of the project was to develop a system for searching phrases in collections of recordings containing speech in Polish. The system accepts queries in text or spoken form, and responds by indicating the audio segment where the speaker uttered the searched phrase. The task was achieved through an innovative combination of speech recognition technology and natural language processing.
The system was developed in response to market demand, related to the increased capability of collecting audio data while lacking tools that enable efficient information retrieval in such data. Audioscope was intended to fill the existing gap in the Polish market, while simultaneously expanding the functionality of products offered by Neurosoft in projects such as “e-Protokół” (recording sessions of common courts) or media monitoring.
Key features of the developed system are:
- it is designed for searching recordings of any speaker speaking Polish (speaker independent) with the ability to adapt to inter-speaker variability,
- the recognized speech is natural speech, with minor restrictions resulting from a certain discipline of speech,
- environmental conditions are not too challenging, but there will certainly be linear transmission distortions, not too much background noise, and slight reverberation,
- it is assumed that there are no significant limitations in computational power, and recognition is not performed in real-time,
- searching in the recognized speech is carried out in real-time, and the system is capable of handling a large collection of documents.
The project implemented in collaboration with the Faculty of Electronics at Wrocław University of Science and Technology and the Faculty of Computer Science at the University
