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

TL;DR — WildElder is a real-world Mandarin Chinese elderly speech dataset containing 33.7 hours across 23,701 manually segmented and annotated utterances from online videos, serving as a challenging benchmark where fine-tuned Conformer-WenetSpeech achieves a 13.54% character error rate.

Key contributions

• Constructed WildElder, an in-the-wild Mandarin elderly speech corpus comprising 33.7 hours of audio extracted from 619 online videos.
• Provided fine-grained manual annotations including orthographic transcripts, speaker age groups, gender, and a standardized 3-tier accent intensity rubric.
• Established comprehensive baseline evaluations covering scratch-trained architectures (Transformer, Conformer, Branchformer, Paraformer) and pre-trained models (Whisper, Conformer-WenetSpeech).
• Analyzed ASR performance variations across demographic axes, revealing significant error rate disparities tied to advanced age brackets and gender.

Problem

Speech recognition technologies struggle with elderly voices due to age-related vocal traits such as reduced volume, slower articulation, and tremors. Existing Mandarin resources like AISHELL-ASR0060, MECSD, and SeniorTalk rely heavily on controlled laboratory or scripted environments, lacking natural spontaneity, topical variety, and acoustic diversity. Consequently, models trained on these datasets often fail in real-world scenarios, highlighting the critical need for a diverse in-the-wild elderly speech benchmark.

Method

The WildElder dataset was assembled via broad keyword searches (e.g., 'nursing home') and targeted channel collection of elderly content creators on online platforms, yielding 71+ raw hours before manual filtering. Human annotators segmented utterances at sentence boundaries, discarding segments with heavy background noise, music, overlap, or non-elderly speakers. Transcripts preserve natural disfluencies while standardizing numbers to uppercase Chinese numerals and alphabetic characters to uppercase ASCII. Annotators also labelled speaker age groups, gender, and accent levels using a qualitative rubric (Light, Moderate, Heavy).

Baselines trained from scratch include Transformer, Conformer (~31.9M params), Branchformer (~29M params), and Paraformer (~31M params), all configured with 32 or 16 batch sizes, 1e-3 learning rates, and trained for 100 epochs using CTC+Attention losses. Pre-trained models evaluated include Conformer-WenetSpeech and Whisper variants (Tiny, Base, Small, Medium), fine-tuned for 20 epochs with learning rates of 4e-5 (CW) and 1e-5 (Whisper) at batch size 16. Decoding experiments contrast CTC greedy, CTC beam search, pure attention, and attention rescoring to balance monotonic alignment stability with contextual linguistic refinement.

Experimental setup

The dataset is partitioned into a 26.7-hour training set (18,835 utterances), a 3.5-hour development set (2,465 utterances), and a 3.5-hour test set (2,400 utterances), all split at the speaker level. Models are evaluated using Character Error Rate (CER, %). Baseline experiments encompass both scratch-trained sequence-to-sequence networks and pre-trained acoustic models.

Results

Models trained from scratch yield high error rates: Conformer achieves the best scratch performance with a 31.74% CER using CTC-Attention with attention rescoring, whereas pure attention decoding struggles severely (e.g., Transformer hits 47.54% CER). Pre-trained models dramatically lower error rates; Conformer-WenetSpeech improves from a 16.43% zero-shot CER to 13.54% after fine-tuning. Whisper models show clear scaling improvements upon fine-tuning, with Whisper-Medium reaching 16.14% CER down from 23.41% zero-shot.

Demographic breakdowns using fine-tuned Conformer-WenetSpeech reveal that female speakers achieve a lower CER of 10.44% compared to 16.89% for male speakers. Furthermore, error rates remain stable between ages 70–85 (~12.8% to 16.2% CER) but sharply degrade beyond age 85, culminating in a 24.41% CER for the 90–95 age bracket.

Limitations

The dataset scope is restricted to Mandarin Chinese, omitting dialectal and cross-lingual variations outside this language. The total duration of 33.7 hours is relatively small compared to web-scale corpora, limiting deep-learning data hunger for multi-billion parameter models. Additionally, representation becomes sparse in extreme age groups above 85 years (e.g., only 30 sentences for ages 90–95), potentially skewing ultra-elderly generalization benchmarks.

Why read this

Researchers building inclusive speech recognition systems and speech-to-text models for elderly populations should read this paper to understand the acoustic hurdles of in-the-wild senior voices. It provides a foundational benchmark dataset and reveals critical performance gaps across demographic age and gender lines.

Code

• https://github.com/NKU-HLT/WildElder

Applications

Voice-controlled assistive interfaces, automated healthcare monitoring systems, and inclusive human-computer interaction tools designed for senior citizens.

Institutions

Nankai University

Funding / 經費: NSF China

Related

• TaigiSpeech: A Low-Resource Real-World Speech Intent Dataset with Scalable Data Mining In-the-Wild — same problem · relatedness 2.0/3
• Hamsa: A Manually Annotated Emirati Arabic Corpus for Speech and Language Technologies — shared data / evaluation · relatedness 2.0/3
• Confidence Score Guided Incremental and Speaker Adaptive Pseudo-Labeling for Semi-Supervised Elderly Speech Recognition — same problem · relatedness 2.0/3
• Decoding while Adapting: Zero-Shot Online Speaker Adaptation via Audio-Textual Prompts for Elderly Speech Recognition — same problem · relatedness 2.0/3
• Imitation Learning for Elder-Facing Speech Synthesis — complementary · relatedness 2.0/3

All 950k paper pairs scored by TypeSafe Jev (scripts/related/); relatedness 0 = unrelated … 3 = directly comparable.