Author ORCID Identifier

0009-0000-5713-2002

Document Type

Conference Paper

Disciplines

1.2 COMPUTER AND INFORMATION SCIENCE, Computer Sciences, Information Science

Publication Details

https://aics2024.ucd.ie/

doi:10.21427/5r8h-px25

Abstract

Conversational avatars provide innovative platforms for enhancing therapist-patient interactions in speech therapy by offering real-time feedback. However, the performance of Automatic Speech Recognition (ASR) models on disordered speech, such as dysarthria and stuttering, remains underexplored. The effectiveness of these systems hinges on the accuracy and processing speed of ASR models when transcribing pathological speech, particularly in real-time scenarios. This study evaluates several pre-trained ASR models, including Whisper-large- v3-turbo, Canary, DistilWhisper, and NVIDIA’s stt-en-fastconformer-ctc-large across three datasets: Common Voice (standard speech), TORGO (dysarthric speech), and UCLASS (stuttered speech). We assess the models using Word Error Rate (WER), Real-Time Factor (RTF), and BERTScore to measure transcription accuracy, computational efficiency, and semantic congruence. The stt-en-fastconformer-ctc-large model demonstrates the fastest processing speeds, achieving the lowest WER and highest BERTScores on both the Common Voice and TORGO datasets, making it highly suitable for real-time therapeutic applications. However, all models struggle with accurately transcribing stuttered speech from the UCLASS dataset. These results highlight the need for ASR improvements for disordered speech, focusing on edge deployment to reduce latency and enhance accuracy with multimodal inputs.

DOI

https://doi.org/10.21427/5r8h-px25

Creative Commons License

Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License
This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License.

Download
Contact the Author


Share

COinS