What is SeniorTalk
SeniorTalk is the world's first Chinese conversational speech dataset for the very elderly, launched by the BAAI Institute in collaboration with the Human Language Technology Lab (HLT Lab) of the College of Computer Science at Nankai University. The dataset contains speech data from 202 very elderly individuals aged 75 and above, with a total duration of 55.53 hours. Data collection covers 16 provinces and cities, encompassing different regional accents, and is based on two-person spontaneous conversations, with topics involving retirement, health, life, etc., closely resembling real communication scenarios. The dataset includes multi-dimensional fine-grained annotations, such as speaker information, conversation content transcription, timestamps (sentence-level and word-level), accent category labels, etc. SeniorTalk provides valuable support for in-depth research on elderly speech signals and optimization of elderly speech interaction systems, promoting the development of related industries such as device age-friendliness, health management, and assistive elderly care robots.
Main functions of SeniorTalk
- Speech recognition: Improve the accuracy of speech recognition for the very elderly, help develop more precise speech recognition systems, and facilitate the use of voice interaction by the elderly.
- Speaker verification: Support research on speaker verification technology to ensure the security and reliability of voice interaction.
- Speaker separation: Provide multi-speaker conversation data, support research on speaker separation technology, and help accurately identify the speech of different speakers in complex environments.
- Speech editing: Provide natural conversation data, support research on speech editing technology, and improve speech synthesis and editing effects.
- Health monitoring and assisted communication: Analyze the speech characteristics of the very elderly, support research on health monitoring and assisted communication technology, and provide data support for elderly care and health management.
Technical principles of SeniorTalk
- Data collection: Based on two-person spontaneous conversations, simulate real communication scenarios to ensure the naturalness and authenticity of speech data. Use a variety of smartphones (including Android and Apple devices) for recording to ensure the diversity and applicability of the data. Follow strict legal and ethical norms to ensure that the data collection process is legal and safe, and protect the privacy of participants.
- Data annotation: Includes speaker information (such as age, gender, region, device, etc.), conversation content transcription, timestamps (sentence-level and word-level), accent category labels, etc. Based on manual annotation and proofreading, ensure the accuracy and completeness of the data.
- Data processing: Based on the 16kHz sampling rate WAV file format, ensure audio quality. Divide the dataset into training, validation, and test sets to support the needs of different research tasks.
- Technical applications: Based on advanced models such as Transformer, Conformer, E-Branchformer for training, improve the performance of speech recognition. Apply models such as X-vector, ResNet-TDNN, ECAPA-TDNN for research on speaker verification and separation. Use methods such as CampNet, EditSpeech, A3T to study speech editing technology and improve the effect of speech synthesis.
Project addresses of SeniorTalk
- GitHub repository：https://github.com/flageval-baai/SeniorTalk
- HuggingFace model library：https://huggingface.co/datasets/BAAI/SeniorTalk
- arXiv technical paper：https://www.arxiv.org/pdf/2503.16578
Application scenarios of SeniorTalk
- Smart elderly care system: Control home appliances and query information based on voice commands, improve the convenience of life for the elderly; monitor voice health status in real time and provide early warnings.
- Assisted communication devices: Help elderly people with language barriers express themselves naturally and accurately recognize specific voice commands in multi-person conversations.
- Health Management Platform: Analyze voice features to assess health conditions and provide voice-interactive health consultation and reminder functions.
- Smart Voice Assistant: Optimize the performance of voice assistants among elderly users and provide more natural and easy-to-understand voice feedback.
- Age-Friendly Product Development: Support the development of smart devices suitable for elderly people and ensure that voice interaction functions adapt to the usage habits of elderly people.