B
B9
Text
Spoken Singlish Corpus and Multimodal Audio-Language Modeling
The construction of the Multitask National Speech Corpus (MNSC), the largest standardized and human-annotated corpus of spoken Singlish (an English-based creole), together with SingAudioLLM, a multitask multimodal audio-language model that jointly handles automatic speech recognition, spoken question answering, spoken dialogue summarization and paralinguistic question answering on that data. The work releases standardized train and test splits and reports state-of-the-art results, providing benchmark resources for spoken Singlish understanding.