WildElder: Mandarin Elderly Speech Corpus
- WildElder is a realistic, manually curated Mandarin Chinese elderly speech corpus that captures spontaneous, in-the-wild audio from online videos.
- The dataset features detailed manual segmentation, orthographic transcription, and metadata annotations such as age, gender, and accent strength to support ASR and speaker profiling.
- Empirical results show robust ASR challenges with higher error rates correlating to age and accent variations, highlighting its significance for real-world deployment.
WildElder is a Mandarin Chinese elderly speech corpus designed as a realistic, manually curated benchmark for speech technology on older speakers. It was created to address a gap in prior Chinese elderly speech resources, which were mostly collected in controlled or laboratory-style settings and therefore offered limited spontaneity, topical diversity, accent diversity, and acoustic variability. By harvesting naturally occurring elderly speech from online videos and enriching it with utterance-level manual segmentation, orthographic transcription, and metadata labels for age group, gender, and accent strength, WildElder is positioned as an “in the wild” resource for automatic speech recognition (ASR) and speaker profiling under conditions closer to real deployment (Wang et al., 10 Oct 2025).
1. Motivation and position within Chinese elderly speech research
WildElder is motivated by the mismatch between existing Chinese elderly speech datasets and the acoustic conditions faced by practical speech systems. The resource is explicitly contrasted with AISHELL-ASR0060, which contains read speech from speakers aged 55+, MECSD, which is aimed at cognitive health research but relies on controlled tasks and recording conditions, and SeniorTalk, which focuses on conversational elderly speech but is still gathered in structured settings with recruited participants. In this framing, earlier datasets remain useful, but they are not sufficient for systems that must operate robustly on noisy, spontaneous, variable speech in everyday environments (Wang et al., 10 Oct 2025).
The “in the wild” property is central rather than incidental. WildElder draws speech from online videos spanning varied topics, recording devices, channels, environments, and speaking styles. This yields naturally occurring spontaneity, diverse acoustic conditions, and regional accent variation. The intended significance is deployment-oriented: voice assistants, health-related speech tools, and general human-computer interaction for older adults must handle variable and non-studio speech rather than only clean, scripted recordings.
A common misconception is that “in the wild” corpora are necessarily weakly supervised or largely auto-labeled. WildElder is presented as an alternative design point: real-world source material is combined with manual segmentation, manual transcription, and manual metadata annotation, with quality checking intended to avoid the annotation unreliability associated with fully automatic web-scale pipelines.
2. Collection pipeline and annotation scheme
The dataset construction pipeline has three stages: data collection, data processing with manual annotation, and quality checking. In the data collection stage, online videos were used as the raw source. Because elderly speech is not systematically available online, two complementary collection strategies were adopted. The first was a keyword-driven search using a custom keyword list; examples given include “80 years old” and “nursing home.” The second was identification of channels featuring elderly content creators, which provided a more efficient route to collecting speech with relatively better reliability and quality. Together these strategies yielded 619 videos totaling more than 71 hours before later filtering and segmentation (Wang et al., 10 Oct 2025).
Preprocessing and segmentation were manual rather than automated. Raw audio-video materials were manually segmented into utterance-level clips based on sentence boundaries and speaker turns. Segments were removed if they involved non-elderly speakers, excessive background music, overlapping speech, or severe distortion. Each retained utterance was transcribed manually into orthographic Chinese characters. Annotators also labeled speaker age group, gender, and accent strength.
The annotation schema is fine-grained but not exhaustively formalized in the visible description. Every utterance has an orthographic Chinese transcription. Age group is estimated from visual and auditory cues rather than verified identity documents. Gender is annotated as speaker metadata. Accent strength is labeled manually using degree labels rather than regional-origin labels, with three categories indicated by the figures and analysis: light, medium, and heavy. The age buckets used in analysis are 70–75, 75–80, 80–85, 85–90, and 90–95.
Quality control is described through four checks: accuracy, completeness, consistency, and usability. “Accuracy” involved sampling transcripts and labels against the audio. “Completeness” ensured that every segment had audio, transcript, and metadata and was free of corruption. “Consistency” checked that annotators followed uniform guidelines. “Usability” verified that file formats and segment boundaries matched the dataset specification. At the same time, WildElder does not report inter-annotator agreement scores, kappa statistics, double annotation, external validation of estimated ages, or detailed adjudication procedures. That omission is consequential for anyone studying metadata reliability or demographic labeling error.
3. Corpus composition, statistics, and partitioning
The final release contains 23,701 utterances and 33.7 hours of audio, derived from the initial 619 videos. The language is Mandarin Chinese, although some utterances carry regional accents. The dataset emphasizes broad topic diversity, including family and daily life, nature and environment, science and education, history and culture, and health and disease. The accompanying word-cloud analysis is used by the authors to argue that the transcripts also cover technology, culture, mathematics, and social issues, which suggests broader semantic coverage than many earlier elderly speech resources (Wang et al., 10 Oct 2025).
Most utterances are short. The median utterance duration is 3.91 seconds, and the median transcript length is 16 Chinese characters. Both duration and text length follow long-tail distributions. Later split-level reporting gives an average utterance duration of about 5.1 seconds across train, development, and test, consistent with a skewed distribution in which the mean exceeds the median.
The split is speaker-level, which is important for preventing speaker leakage in ASR evaluation.
| Subset | Utterances | Audio hours |
|---|---|---|
| Train | 18,835 | 26.7 |
| Development | 2,465 | 3.5 |
| Test | 2,400 | 3.5 |
The publication does not state the number of distinct speakers in the full corpus or in each split. It also does not provide exact corpus-wide totals for gender balance, accent-category distribution, or age-group distribution. Instead, it offers qualitative summaries: the majority of utterances come from speakers aged 70 to 85; speech from people above 90 is included in smaller quantity; both male and female speakers contribute substantially; and light and medium accent levels dominate, with fewer heavy-accent utterances.
Some demographic detail is available for the 2,400-sentence test set. Female speech accounts for 1,282 sentences and male speech for 1,118. Age-group counts are 1,477 for 70–75, 383 for 75–80, 380 for 80–85, 130 for 85–90, and 30 for 90–95. These counts are specific to the test set and are not stated to be identical to corpus-wide totals.
4. Benchmark formulation and modeled tasks
WildElder is framed as supporting both ASR and speaker profiling because every utterance carries age-group, gender, and accent-strength labels. In the visible experiments, however, the benchmarking is ASR-centric. The paper does not report standalone classification experiments for age-group prediction, gender classification, or accent-strength prediction. A plausible implication is that WildElder’s speaker-profiling value is presently infrastructural: it enables such work, but does not yet define standard profiling baselines (Wang et al., 10 Oct 2025).
The ASR baselines are divided into two groups. The scratch-trained group contains Transformer, Conformer, Branchformer, and Paraformer. Transformer is described as a pure self-attention sequence-to-sequence architecture. Conformer augments Transformer-style modeling with convolution to better capture local acoustic cues. Branchformer uses a branched attention–feedforward design. Paraformer is a non-autoregressive ASR model combining CTC and parallel decoding for faster inference. The pre-trained group consists of Conformer-WenetSpeech and Whisper models of sizes Tiny, Base, Small, and Medium.
Training configurations are explicitly summarized. For scratch training, Transformer uses batch size 32, learning rate , and 100 epochs; Conformer uses batch size 32, learning rate , and 100 epochs; Branchformer uses batch size 16, learning rate , and 100 epochs; and Paraformer uses batch size 32, learning rate , and 100 epochs. For pre-trained models, Conformer-WenetSpeech uses batch size 16, learning rate , and 20 epochs, while Whisper uses batch size 16, learning rate , and 20 epochs. Evaluation uses character error rate (CER), which is appropriate for Chinese ASR.
The visible description does not specify acoustic feature extraction, waveform front ends, sampling rate, tokenizer design, or text normalization. It also does not provide explicit loss equations or metric formulas. The reported loss labels are “CTC+ATT” for Transformer, Conformer, and Branchformer, and “CTC+Paraformer” for Paraformer, but no formal mathematical definitions are included.
5. ASR baselines, decoding results, and empirical difficulty
WildElder is intentionally difficult as an ASR benchmark. The scratch-trained results are reported under multiple decoding strategies: CTC greedy, CTC beam, attention, and attention rescoring. Transformer with 29.80M parameters and loss CTC+ATT yields CERs of 37.44, 37.28, 47.54, and 36.30 under those four strategies, respectively. Conformer with 31.94M parameters and loss CTC+ATT yields 32.51, 32.48, 39.58, and 31.74. Branchformer with 29.01M parameters and loss CTC+ATT yields 35.67, 35.61, 45.60, and 34.80. Paraformer with 31.04M parameters and loss CTC+Paraformer yields 42.71 for CTC greedy and 38.39 for CTC beam, with the attention-based columns not applicable (Wang et al., 10 Oct 2025).
Several patterns follow directly from these results. First, all from-scratch systems perform relatively poorly, with CERs mostly in the 30–40% range. Second, Conformer is the strongest scratch-trained architecture. Third, pure attention decoding is markedly worse than CTC-based decoding on this dataset. Fourth, attention rescoring improves over pure attention decoding and, for Conformer, also improves over plain CTC beam.
Pre-trained models are substantially stronger but remain far from saturated. Conformer-WenetSpeech reports 16.43 zero-shot CER and 13.54 after fine-tuning. Whisper-Tiny reports 53.50 zero-shot and 32.20 after fine-tuning. Whisper-Base reports 41.67 and 26.61. Whisper-Small reports 29.00 and 20.25. Whisper-Medium reports 23.41 and 16.14. These numbers show that domain adaptation matters strongly and that Conformer-WenetSpeech is the best reported system, both zero-shot and after fine-tuning, with a best CER of 13.54. Within Whisper, larger models outperform smaller ones in both regimes.
The benchmark therefore occupies an intermediate regime between narrow, clean laboratory corpora and much larger but less targeted web-scale speech resources. This suggests that WildElder’s difficulty arises from a combination of elderly-specific speech characteristics and in-the-wild acoustic variation rather than from scale alone.
6. Demographic effects, interpretive value, and stated limitations
WildElder includes subgroup analysis for the fine-tuned Conformer-WenetSpeech system. On gender, female speech is substantially easier than male speech: female CER is 10.44%, while male CER is 16.89%. The error-type breakdown reports, for female speech, substitution 7.12%, deletion 2.13%, and insertion 1.19%; for male speech, substitution 10.77%, deletion 3.18%, and insertion 2.94%. On age, reported CERs are 12.85 for ages 70–75, 16.17 for 75–80, 12.77 for 80–85, 20.06 for 85–90, and 24.41 for 90–95. The corresponding test-set sentence counts are 1,477, 383, 380, 130, and 30, respectively. The authors interpret these results as evidence that recognition generally becomes harder with more advanced age, with especially sharp degradation beyond age 85 (Wang et al., 10 Oct 2025).
The dataset’s interpretive significance lies in the conjunction of three properties: real-world acoustic diversity, elderly-specific focus, and fine-grained manual annotation. This combination makes it suitable for robust ASR on older speakers, domain adaptation from large pre-trained models, subgroup robustness and fairness analysis across age and gender, metadata-aware or multi-task modeling, accent-robust recognition, and future speaker-profiling work. Because accent strength is labeled as light, medium, or heavy rather than by province or dialect type, it is particularly suited to robustness studies indexed by accent intensity rather than to fine-grained regional linguistic analysis.
Its limitations are also explicit. Age is estimated rather than verified, introducing unavoidable label uncertainty. The number of unique speakers is not reported. Accent origin is not specified. The quality-control process is described conceptually, but no inter-annotator agreement or annotation error rates are provided. Recording-condition metadata such as indoor versus outdoor environment, microphone type, reverberation class, or noise category are not explicitly annotated. Despite positioning the dataset as useful for speaker profiling, the visible experimental section lacks dedicated profiling baselines. The corpus size, 33.7 hours, is modest relative to very large ASR corpora, although this is partly offset by careful manual curation and its narrow elderly-speech focus. The visible text also reports no ablation studies, no detailed qualitative error analysis beyond subgroup CER tables, and no explicit ethical, privacy, consent, or licensing discussion.
WildElder is therefore best understood as a challenging benchmark rather than a closed problem. It documents the gap between elderly-oriented deployment conditions and the performance of contemporary ASR models, while providing a manually curated corpus that makes that gap measurable.
