CoChat

Speech Emotion Recognition via an Attentive Time–Frequency Neural Network

Open access
Lu, Cheng, Zheng, Wenming, Lian, Hailun, Zong, Yuan, Tang, Chuangao, Li, Sunan, Zhao, Yan · 2023 · IEEE Transactions on Computational Social Systems

The DOI matches its Crossref record; 38 of the 38 references it cites were found in a registry or an index; 3 of the 28 sources its claims cite were checked for support.

Abstract

Spectrogram is commonly used as the input feature of deep neural networks to learn the high(er)-level time–frequency pattern of speech signal for speech emotion recognition (SER). Generally, different emotions correspond to specific energy activations both within frequency bands and time frames on spectrogram, which indicates the frequency and time domains are both essential to represent the emotion for SER. However, recent spectrogram-based works mainly focus on modeling the long-term dependency in time domain, which makes these methods suffer from the following issues: 1) neglecting to model the emotion-related correlations within frequency domain during the time–frequency joint learning and 2) ignoring to capture the specific frequency bands associated with emotions. To cope with the issues, we propose an attentive time–frequency neural network (ATFNN) for SER, including a time–frequency neural network (TFNN) and time–frequency attention. Specifically, aiming at the first issue, we design a TFNN with a frequency-domain encoder (F-Encoder) based on the Transformer encoder and a time-domain encoder (T-Encoder) based on the bidirectional long short-term memory (Bi-LSTM). The F-Encoder and T-Encoder model the correlations within frequency bands and time frames, respectively, and they are embedded into a time–frequency joint learning strategy to obtain the time–frequency patterns of speech emotions. Moreover, to handle the second issue, we adopt the time–frequency attention with a frequency-attention network (F-Attention) and a time-attention network (T-Attention) to focus on the emotion-related long-range dependencies between frequency bands and across time frames, which can enhance the emotional discrimination of speech features. Extensive experimental results on three public emotional databases, i.e., IEMOCAP, ABC, and CASIA, show that our proposed ATFNN outperforms the state-of-the-art methods.

Overview

CoChat Summary

TL;DR

The paper proposes an attentive time–frequency neural network (ATFNN) for speech emotion recognition, combining frequency-domain and time-domain encoders with frequency and time attention. It reports state-of-the-art results on IEMOCAP, ABC, and CASIA.

Key findings

  • ATFNN achieved the best WAR on IEMOCAP at 73.81%, with UAR 64.48%; ADARL had higher UAR at 65.86%.
  • On ABC, ATFNN obtained WAR 68.84% and UAR 57.57%, outperforming LLDs + HMM/GMM, RGSR-LPDA, LLDs + SVM, and GerDA.
  • On CASIA, ATFNN reached WAR and UAR of 48.75%, higher than LLD + DR, DNN + ELM, HuWSF, and DTPM.
  • Ablation showed ATFNN (73.81% WAR on IEMOCAP) outperformed TFNN (72.99%), ATNN (71.76%), and AFNN (70.71%); attention visualizations indicated happy has increased high-frequency energy and sad decreased high-frequency energy.

Methods

  • The ATFNN embeds time–frequency attention into a time–frequency neural network: an F-Encoder based on a Transformer encoder without position encoding, a T-Encoder based on a three-layer Bi-LSTM with 128 hidden nodes, plus F-Attention and T-Attention convolution blocks.
  • Experiments used IEMOCAP (2,280 utterances, 4 emotions), ABC (430 utterances, 6 emotions), and CASIA (1,200 utterances, 6 emotions), with leave-one-speaker-out cross-validation and WAR/UAR metrics.
  • Inputs were log-Mel-spectrograms extracted by STFT with 20 ms Hamming window, 50% overlap, 512-point FFT, and 80 Mel bands; utterances were divided into 128-frame segments. The model was trained from scratch in PyTorch on RTX3090 GPUs with batch size 128 and Adam learning rate 0.0005.

Limitations

  • The paper states that precise correspondence between different emotions and frequency bands should be deeply explored in future research.
  • The paper notes that IEMOCAP and ABC are class-imbalanced, causing WAR/UAR discrepancy and confusion, e.g., 58.27% of happy samples recognized as neutral in IEMOCAP and 45.45% of intoxicate samples recognized as cheerful in ABC.
  • The paper states CASIA performance is much lower than IEMOCAP and ABC, possibly because CASIA forces speakers to express different emotions in the same sentence and is a tonal language; it also notes ADARL’s domain adaptation gives better UAR on IEMOCAP than ATFNN.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tcss.2022.3219825

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Verification

Claims the paper makes

1 Valid · 0 Not valid · 13 Not checked · 46 Own results and interpretations

30 more claims

Create a free account to see all 60 claims and to verify the 25 references still open.

See all claims — free account
  • It may state two separate claims as one. 18
  • Probably not a substantial claim. 9
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

5 of 5

Citations

Cites

Show all 38 references

Research works printed but not linked: 2

  1. Zhang, Design of speech corpus for Mandarin text to speech, 1
  2. 10.1037/0003-066x.50.5.372 · DOI · Paper page

Cited by

See all 51

Metadata

Authors
Lu, C., Zheng, W., Lian, H., Zong, Y., Tang, C., Li, S., & Zhao, Y.
Affiliations
Southeast University (CN)
Venue
IEEE Transactions on Computational Social Systems
Published
2023-12
Volume
10
Issue
6
Pages
3159-3168
ISSN
  • 2329-924X (ISSN-L)
  • 2373-7476
OpenAlex
W4312857458
Cited by
51
References
38
Citations by year
2026: 12 · 2025: 22 · 2024: 10 · 2023: 6
FWCI
8.98
Type
journal article
Language
en
Open access
Locations
  • IEEE Transactions on Computational Social Systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • NSFC (U2003207, 61921004, 61902064)
  • Jiangsu Frontier Technology Basic Research Project (BK20192004)
  • Zhishan Young Scholarship of Southeast University
  • Scientific Research Foundation of Graduate School of Southeast University (YBPY1955)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Emotion and Mood Recognition
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
  • Physical Sciences › Computer Science › Signal Processing › Speech and Audio Processing
Keywords
speech emotion recognition, spectrogram, time-frequency attention, Transformer encoder, bidirectional LSTM, long-range dependencies

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-28, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-09-28 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Speech Emotion Recognition via an Attentive Time–Frequency Neural Network. https://app.cochat.ai/papers/10.1109/tcss.2022.3219825
The paper
Lu, C., Zheng, W., Lian, H., Zong, Y., Tang, C., Li, S., & Zhao, Y. (2023). Speech Emotion Recognition via an Attentive Time–Frequency Neural Network. IEEE Transactions on Computational Social Systems. https://doi.org/10.1109/tcss.2022.3219825

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.