Research map / Large language models / Language, safety and interpretability
SeniorTalk: A Chinese Conversation Dataset with Rich Annotations for Super-Aged Seniors
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SeniorTalk, a 55.53-hour Mandarin spontaneous conversation dataset from 202 speakers aged 75+ with multi-level annotations, and benchmarks speaker verification, diarization, ASR, and speech editing on it.
Problem
Existing elderly speech corpora under-represent super-aged (75+) speakers, rely on reading-style speech, and lack rich annotations (diarization, dialect/accent, paralinguistic markers), hindering voice technologies for aging populations.
Approach
A conversational Mandarin dataset recorded from super-aged speakers (75-85) across 16 provinces with speaker metadata, session-level diarization and overlap labels, utterance-level transcripts with accent intensity, and token-level paralinguistic markers, plus benchmark experiments on four speech tasks.
Key results
- ECAPA-TDNN achieves the lowest test EER on SeniorTalk speaker verification (8.56% PLDA, 5.66% cosine), outperforming x-vector and ResNet-TDNN.
- Fine-tuned Whisper-large-v3 CER drops from 57.74% zero-shot to 23.84%, while fine-tuned Paraformer-large reaches 14.41% CER.
- Among models trained from scratch, E-Branchformer achieves 33.25% CER, outperforming Conformer by 1.36% and Transformer by 15.74%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune ECAPA-TDNN on SeniorTalk | ECAPA-TDNN speaker verification · Speaker verification test | EER (cosine similarity, %) | Helped |
| replace the default ResNet-34-LM extractor with ECAPA-TDNN fine-tuned on SeniorTalk | PyAnnote speaker diarization · Speaker diarization, collar=0 | DER (%): 33.14 -> 27.84; 5.3% DER improvement | Helped |
| replace the default ResNet-34-LM extractor with ECAPA-TDNN fine-tuned on SeniorTalk | PyAnnote speaker diarization · Speaker diarization, collar=0.25 | DER (%): 28.39 -> 22.85; 5.54% DER improvement | Helped |
| use E-Branchformer instead of Conformer | ASR models trained from scratch · ASR test | CER (%): 34.61 -> 33.25; 1.36% | Helped |
| fine-tune Whisper-large-v3 on SeniorTalk | Whisper-large-v3 ASR · ASR test | CER (%): 57.74 -> 23.84 | Helped |
| fine-tune Paraformer-large on SeniorTalk | Paraformer-large ASR · ASR test | CER (%): 14.91 -> 14.41 | Helped |
| increase accent intensity | E-Branchformer, Conformer, and Transformer ASR · ASR test accent groups | CER (%) | Hurt |
| use southern rather than northern regional speech | E-Branchformer ASR · ASR test regions | CER (%): 0.97% | Helped |
| use FluentSpeech instead of other evaluated speech-editing models | Speech editing · Speech editing test | MOS | Helped |
Surprising findings
- Zero-shot Whisper models exceed 50% CER on elderly conversational speech, with the smallest-to-medium variants performing worse than expected.
- ECAPA-TDNN wins on the test set but underperforms ResNet-TDNN on the development set, suggesting overfitting risk.
Limitations
- Gender imbalance (67 male vs 135 female) and regional/accent biases in the speaker pool.
- Age range covers 75-85 only and does not reach the 85+ population; dataset is relatively small (55.53 hours).
- Dataset access restricted to verified academic researchers under CC BY-NC-SA 4.0 due to voice-synthesis misuse risk.
Experimental setup
Datasets: SeniorTalk VoxCeleb
Benchmarks: SeniorTalk
Metrics: EER minDCF DER CER MCD STOI PESQ MOS Boundary-MOS
Baselines: x-vector ECAPA-TDNN ResNet-TDNN ResNet-34-LM Transformer Conformer E-Branchformer Paraformer-large Whisper CampNet EditSpeech A3T
Builds on: WeNet SpeechBrain PyAnnote Whisper-Flamingo Speech-Editing-Toolkit
Closest papers by meaning
- MultiTalk: Scaling Full-Duplex Speech Models to Long, Multi-Party, Bilingual Conversation (NeurIPS 2026 · similarity 0.85)
- SD-Eval: A Benchmark Dataset for Spoken Dialogue Understanding Beyond Words (NeurIPS 2024 · similarity 0.83)
- IndicVoices-R: Unlocking a Massive Multilingual Multi-speaker Speech Corpus for Scaling Indian TTS (NeurIPS 2024 · similarity 0.82)
- Aligning Spoken Dialogue Models from User Interactions (ICML 2025 · similarity 0.82)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.82)
- TalkCuts: A Large-Scale Dataset for Multi-Shot Human Speech Video Generation (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.