Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WildElder: A Chinese Elderly Speech Dataset from the Wild with Fine-Grained Manual Annotations

Hui Wang, Jiaming Zhou, Jiabei He, Haoqin Sun, Yong Qin

TL;DR

WildElder targets the gap in Chinese elderly speech resources by sourcing in-the-wild Mandarin audio from online videos and applying fine-grained manual annotations (age group, gender, accent strength). It describes a three-stage construction pipeline (collection, annotation, quality control) and provides comprehensive statistics, including 23,701 utterances (33.7 hours) from 619 videos with diverse topics. The paper evaluates both scratch-trained and pre-trained models, showing that elderly speech remains difficult but gains are achievable through domain adaptation, with Conformer-WenetSpeech and larger Whisper models performing best after fine-tuning. Collectively, WildElder offers a challenging, demographically rich benchmark for robust elderly ASR and speaker profiling, enabling demographic-aware modeling and future research in inclusive HCI.

Abstract

Elderly speech poses unique challenges for automatic processing due to age-related changes such as slower articulation and vocal tremors. Existing Chinese datasets are mostly recorded in controlled environments, limiting their diversity and real-world applicability. To address this gap, we present WildElder, a Mandarin elderly speech corpus collected from online videos and enriched with fine-grained manual annotations, including transcription, speaker age, gender, and accent strength. Combining the realism of in-the-wild data with expert curation, WildElder enables robust research on automatic speech recognition and speaker profiling. Experimental results reveal both the difficulties of elderly speech recognition and the potential of WildElder as a challenging new benchmark. The dataset and code are available at https://github.com/NKU-HLT/WildElder.

WildElder: A Chinese Elderly Speech Dataset from the Wild with Fine-Grained Manual Annotations

TL;DR

WildElder targets the gap in Chinese elderly speech resources by sourcing in-the-wild Mandarin audio from online videos and applying fine-grained manual annotations (age group, gender, accent strength). It describes a three-stage construction pipeline (collection, annotation, quality control) and provides comprehensive statistics, including 23,701 utterances (33.7 hours) from 619 videos with diverse topics. The paper evaluates both scratch-trained and pre-trained models, showing that elderly speech remains difficult but gains are achievable through domain adaptation, with Conformer-WenetSpeech and larger Whisper models performing best after fine-tuning. Collectively, WildElder offers a challenging, demographically rich benchmark for robust elderly ASR and speaker profiling, enabling demographic-aware modeling and future research in inclusive HCI.

Abstract

Elderly speech poses unique challenges for automatic processing due to age-related changes such as slower articulation and vocal tremors. Existing Chinese datasets are mostly recorded in controlled environments, limiting their diversity and real-world applicability. To address this gap, we present WildElder, a Mandarin elderly speech corpus collected from online videos and enriched with fine-grained manual annotations, including transcription, speaker age, gender, and accent strength. Combining the realism of in-the-wild data with expert curation, WildElder enables robust research on automatic speech recognition and speaker profiling. Experimental results reveal both the difficulties of elderly speech recognition and the potential of WildElder as a challenging new benchmark. The dataset and code are available at https://github.com/NKU-HLT/WildElder.

Paper Structure

This paper contains 19 sections, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Dataset Description
  3. Construction Process
  4. Data Collection
  5. Data Processing and Annotation
  6. Quality Check
  7. Dataset Statistics
  8. Demographic Diversity
  9. Acoustic and Linguistic Richness
  10. Experiment Setup
  11. Dataset Splits
  12. Baseline Models
  13. Models trained from scratch
  14. Pre-trained models.
  15. Results and Analysis
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Workflow of the WildElder dataset construction, including data collection, annotation, and quality check.
  • Figure 2: Utterance-level distribution across age groups, separated by gender (male on the left, female on the right) and accent strength (light, medium, heavy).
  • Figure 3: Distribution of utterance duration and character count.
  • Figure 4: Word cloud of the dataset transcripts, showing frequently used characters and topics.