Automated Processing of Speech Recordings for Dietary Assessment: Evaluation in the LLMIC Context

Novel methods of using smartphones to collect speech records of dietary intake facilitate self-reporting of intake in free-living conditions, and may further reduce biases affecting reliable capture such as complexity and reactivity. However, the processing of speech records for dietary assessment is a time-consuming task for analysts and thus restricted by research infrastructure. Low- and Low-Middle Income Countries have been disproportionately affected by barriers to dietary assessment (e.g., due to varying literacy or limited research infrastructure). As such, speech records could be a promising avenue to address the research gap in Low and Low-Middle Income regions. While recent advances in speech recognition and natural language processing technology have facilitated automation of this process, no existing studies have evaluated the effectiveness of this technology when applied to an LLMIC context. To this end we adapt the methods identified in a review of the literature to this new context and evaluate their performance on a data set of Khmer speech recordings describing dietary data captured in a free-living context in Cambodia.