Sequence-to-Sequence Models in Italian Atypical Speech Recognition

In the domain of automatic speech recognition (ASR), we explore the usage of a state-of-the-art transformer-based sequence-to-sequence model to build a speaker-dependent isolated word recognizer for native Italian speakers with a speech disorder, such as dysarthria. In particular, this paper is concerned with a self-supervised learning approach, where the Wav2Vec2 has been fine-tuned on our private Italian corpus containing a total of 41 hours of speech contributions authored by 191 individuals with a disability and atypical speech. The discussed approach has been also evaluated thanks to collaboration of sixteen speakers with diverse degrees of speech disorders (mild, moderate, severe), and our analysis has shown a remarkable performance of our ASR system, with an overall word recognition accuracy of 97.4%.