Research map / Large language models / Language, safety and interpretability
Dawn of the Transformer Era in Speech Emotion Recognition: Closing the Valence Gap
Johannes Wagner, Andreas Triantafyllopoulos, Hagen Wierstorf, Maximilian Schmitt, Felix Burkhardt, Florian Eyben, Björn Wolfgang Schuller
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune the transformer layers instead of freezing them (ablation) | wav2vec2 / HuBERT SER models · MSP-Podcast, IEMOCAP, MOSI | CCC | Improved |
| use pre-trained weights instead of random initialisation (ablation) | wav2vec2-large · valence / sentiment, in-domain and cross-corpus | CCC | Improved |
| use pre-trained weights instead of random initialisation (ablation) | wav2vec2-large · arousal and dominance | CCC | Improved |
| fuse explicit BERT text embeddings with fine-tuned audio embeddings | fine-tuned wav2vec2 / HuBERT · MSP-Podcast, IEMOCAP, MOSI | CCC | Mixed |
| pre-train on more, multilingual data (voxpopuli 100k h to xls-r 463k h) | wav2vec2 large · MSP-Podcast, IEMOCAP, MOSI | CCC: 100k and 463k vs 63k hours | Hurt |
| use large architecture instead of base | wav2vec2 · MSP-Podcast, IEMOCAP, MOSI | CCC | No clear effect |
| use HuBERT instead of wav2vec2 (base) | base architecture · MSP-Podcast, IEMOCAP, MOSI valence | valence CCC | Improved |
| remove the top 12 of 24 transformer layers before fine-tuning (ablation) | wav2vec2-large-robust · MSP-Podcast, IEMOCAP, MOSI | CCC: 24 -> 12 layers | No clear effect |
| keep 10 or fewer transformer layers (ablation) | wav2vec2-large-robust · IEMOCAP and MOSI valence | valence / sentiment CCC: 10 or fewer layers | Hurt |
| reduce fine-tuning data to 25% of MSP-Podcast training set (ablation) | pruned wav2vec2-large-robust · arousal and dominance on MSP-Podcast | CCC: loss of less than .02 at 25% | No clear effect |
| use the full training set rather than a reduced fraction (ablation) | pruned wav2vec2-large-robust · valence on MSP-Podcast and MOSI | valence CCC: boost of almost .1 on MOSI for the remaining 25% | Improved |
| add background noise at 0 dB instead of 20 dB SNR | wav2vec2-large-robust · MSP-Podcast valence | change in CCC: -.221 vs -.003 (natural sounds) | Hurt |
Closest papers by meaning
- Dynamic-SUPERB Phase-2: A Collaboratively Expanding Benchmark for Measuring the Capabilities of Spoken Language Models with 180 Tasks (ICLR 2025 · similarity 0.87)
- VowelPrompt: Hearing Speech Emotions from Text via Vowel-level Prosodic Augmentation (ICLR 2026 · similarity 0.86)
- MSE-Adapter: A Lightweight Plugin Endowing LLMs with the Capability to Perform Multimodal Sentiment Analysis and Emotion Recognition (AAAI 2025 · similarity 0.86)
- EmotionThinker: Prosody-Aware Reinforcement Learning for Explainable Speech Emotion Reasoning (ICLR 2026 · similarity 0.86)
- AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models (ICML 2025 · similarity 0.85)
- Emotion-LLaMA: Multimodal Emotion Recognition and Reasoning with Instruction Tuning (NeurIPS 2024 · similarity 0.85)
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.
