TL;DR
This paper introduces the Mandarin Elderly Cognitive Speech Database (MECSD), a novel, large-scale dataset designed to facilitate research into early-stage dementia detection using speech. They demonstrate its utility with a pilot Automatic Speech Recognition (ASR) experiment, showing that speech impairments are detectable even in Mild Cognitive Impairment (MCI) patients.
What Problem Does This Paper Address?
Detecting neurodegenerative diseases like Alzheimer's (AD) and its precursor, Mild Cognitive Impairment (MCI), at their earliest stages is crucial for effective intervention and treatment. While speech and language alterations are known early indicators, the lack of high-quality, specialized speech databases has hindered research and the development of practical screening tools. Existing datasets often suffer from limitations such as small sample sizes, unmatched demographic features, or a focus on moderate to severe dementia stages, making them less suitable for identifying subtle changes associated with MCI. This paper aims to bridge this gap by creating a comprehensive Mandarin Chinese speech database specifically for studying early-stage cognitive decline.
Methods
The core contribution is the design and creation of the Mandarin Elderly Cognitive Speech Database (MECSD). This database was meticulously constructed to capture a range of speech and language behaviors. It includes recordings from 85 native Mandarin Chinese speakers (55-85 years old), comprising 20 individuals diagnosed with MCI and 65 healthy controls. To elicit diverse speech patterns, nine distinct tasks were employed, encompassing conventional cognitive tests (like verbal fluency and naming) and various connected speech productions (e.g., self-introduction, picture description, poem reading). The recordings were made in a controlled, noise-attenuated environment with professional equipment. For analysis, the audio data was segmented at the sentence level, and manual transcriptions were created using the TalkBank CHAT protocol, annotating not only words but also paralinguistic phenomena like filled pauses and disfluencies.
To demonstrate the database's application, the authors conducted a pilot Automatic Speech Recognition (ASR) experiment. They trained four different ASR systems (GMM-HMM, GMM-HMM-MLLR, Hybrid DNN, and Hybrid DNN-SAT) using existing large Mandarin speech corpora and then evaluated them on the MECSD. This experiment aimed to reveal differences in speech production between MCI patients and healthy controls.
Datasets
- Mandarin Elderly Cognitive Speech Database (MECSD): This is the newly introduced benchmark dataset. It contains 110 hours of speech recordings from 85 native Mandarin Chinese speakers (20 MCI, 65 controls), aged 55-85. The database includes manually transcribed materials with temporal information and covers nine different speech tasks.
- Training Data for ASR Systems: The paper mentions using existing large Mandarin speech corpora for training the ASR models. Specifically, they cite the DARPA GALE Mandarin broadcast news data (562 hours), Chinese transcription, TDT4 Mandarin Chinese transcription, and HGSD prompts.
Key Insights and Ideas
- Speech as an Early Biomarker: The study reinforces the idea that subtle changes in speech production can be indicative of early-stage cognitive decline, even before significant clinical symptoms manifest.
- The Need for Specialized Datasets: The creation of MECSD highlights the critical need for well-designed, culturally specific, and age-appropriate speech databases to advance research in this area.
- ASR Sensitivity to Cognitive Impairment: The pilot ASR experiment demonstrates that even with standard ASR techniques, differences in speech characteristics between MCI patients and healthy controls can be detected, suggesting the potential for automated screening.
- Task Diversity is Key: Employing a variety of tasks, from structured repetitions to spontaneous speech, is crucial for capturing a comprehensive picture of speech and language abilities and their potential degradation.
Key Results
The pilot ASR experiment showed that the ASR systems performed worse on the MCI subset compared to the normal subset, as indicated by higher Character Error Rates (CER). For instance, the Hybrid DNN-SAT system achieved an overall CER of 21.95% for the normal subset and 37.70% for the MCI subset. This difference suggests that patients with MCI exhibit speech production impairments that are detectable by ASR technology. The paper also notes that speaker adaptation techniques (MLLR and SAT) significantly improved ASR performance by reducing variability introduced by speakers.
Limitations and Future Work
- Limitations:
- The pilot ASR experiment used only speech data from the Picture Description (PD) and Self-Introduction (SI) tasks, as these were continuous speech data. Other tasks were not utilized in this specific experiment.
- The paper acknowledges that the MCI participants were recruited from a hospital setting, which might introduce some biases.
- Future Work:
- The MECSD is intended to facilitate hypothesis testing and further research into the speech features associated with early-stage dementia.
- The authors suggest that the database can be applied to various areas, including speech signal analysis, speaker recognition, and speech recognition for dementia detection.
- Further research can explore multidimensional linguistic features (lexical, semantic, syntactic, acoustic) at the task level, beyond sentence-level ASR.
Summary updated