Enhancing automatic speech recognition detection of suicidal-related terms in Thai language

Abstract

Suicide and Self-harm are significant public health concerns in Thailand. Many individuals with suicidal risk are still unable to receive timely support. The Mental health hotline 1323 is an important service channel but faces a high volume of incoming calls that exceeds staff capacity. Currently, an automated severity triage service and an automatic speech recognition (ASR) system are used during wait times. However, current ASR models lack accuracy in transcribing domain-specific suicidal terms. This study aims to develop an ASR system for Thai language that can better detect suicidal-related terms from spoken conversations. ASR models were fully fine-tuned using a domain-specific dataset and evaluated by Word Error Rate (WER) and Term-level recall for suicidal-related keywords. Experimental results show that domain fine-tuning significantly improves the detection of suicidal-related terms. The Micro Recall increased from 0.389 to 0.903, and the Word Error Rate (WER) decreased from 57.725 to 25.541. These results suggest that the proposed approach can improve the ability of ASR systems to identify suicidal-related expressions and has potential to support risk screening and prioritization in mental health service systems.