W2v Bert Beto Multihead Emotion En

The fusion mechanism uses cross-attention layers to model interactions between acoustic and textual representations, enabling the system to learn relationships between speech and linguistic emotional cues.

It is designed to jointly exploit complementary emotional information from speech and text in order to improve emotion recognition performance compared to unimodal approaches.

The model predicts one of the following emotion labels:

• angry
• disgust
• fear
• happy
• neutral
• sad
• surprise

Intended use

This model is intended for research and applied scenarios involving multimodal emotion recognition in English, such as:

• multimodal conversational analysis
• speech and text emotion analysis
• affective computing research
• emotion-aware conversational systems
• human-computer interaction
• multimodal AI research

The model is particularly useful in scenarios where both speech audio and transcribed text are available.

It can be used through the speech-emotion toolkit.

Out-of-scope use

This model should not be used as the sole basis for high-stakes decisions, including but not limited to:

• clinical diagnosis
• mental health assessment
• employment, legal, or educational decisions
• biometric profiling or surveillance
• automated decisions affecting individuals without human oversight

Emotion recognition is inherently uncertain and context-dependent. Predictions should be interpreted as model estimates, not as definitive assessments of a person's emotional state.

Training data

The model was trained on the English multimodal datasets used in the speech-emotion project.

The training data combines multiple publicly available English speech and multimodal emotion recognition datasets, including:

• RAVDESS
• TESS
• MELD
• datasets derived from prior speech emotion recognition research benchmarks

Because the original datasets use different emotion taxonomies, all datasets were harmonized into a unified seven-class emotion taxonomy:

• angry
• disgust
• fear
• happy
• neutral
• sad
• surprise

For the English multimodal emotion recognition setup, the same aligned speech-text samples were used for both the acoustic and textual modalities:

• Training samples: 3,622
• Validation samples: 453
• Test samples: 453

More details about the dataset preprocessing and label harmonization pipeline are available in the project repository:

https://github.com/NLP-UMUTeam/umuteam-speech-emotion

Evaluation

The model was evaluated on the English held-out test set used in the speech-emotion toolkit.

Performance comparison on English emotion recognition

The results show that combining acoustic and linguistic representations improves emotion recognition performance compared to unimodal approaches.

The multi-head cross-attention strategy enables interaction between modalities through attention mechanisms and achieves strong multimodal performance for English emotion recognition.

How to use

Install the toolkit:

pip install speech-emotion

Multimodal emotion recognition using audio and text

from speech_emotion import predict_emotion

emotion = predict_emotion(
    audio_path="audio.wav",
    text="I was really happy to see you again.",
    language="en",
    mode="multihead",
    model_config_path="model.json"
)

print("Detected emotion:", emotion)

Multimodal emotion recognition using automatic transcription (Whisper)

If no transcription is provided, the toolkit can automatically generate it using Whisper before performing emotion recognition.

from speech_emotion import predict_emotion

emotion = predict_emotion(
    audio_path="audio.wav",
    language="en",
    mode="multihead",
    model_config_path="model.json"
)

print("Detected emotion:", emotion)

Repository:

https://github.com/NLP-UMUTeam/umuteam-speech-emotion

Limitations

• The model is designed for English multimodal emotion recognition and may not generalize reliably to other languages.
• It predicts a single label from a fixed set of seven emotions.
• Emotion expression is subjective and highly context-dependent.
• Performance may decrease with noisy audio, inaccurate transcriptions, overlapping speakers, or domain shifts.
• The model assumes that audio and text inputs are semantically aligned.
• Errors in automatic speech transcription may negatively affect multimodal performance.
• Attention-based fusion mechanisms may require larger datasets or different optimization strategies to generalize effectively.

Bias and ethical considerations

Emotion recognition systems may reflect biases present in their training data, including differences related to accents, speaking styles, demographics, recording conditions, or annotation subjectivity.

Users should avoid interpreting predictions as objective truths about a person's internal emotional state. The model should be used with transparency, appropriate consent, and human oversight, especially in sensitive contexts.

Citation

If you use this model in your research, please cite the following works:

speech-emotion toolkit

@article{PAN2026102677,
title = {speech-emotion: A multilingual and multimodal toolkit for emotion recognition from speech},
journal = {SoftwareX},
volume = {34},
pages = {102677},
year = {2026},
issn = {2352-7110},
doi = {https://doi.org/10.1016/j.softx.2026.102677},
url = {https://www.sciencedirect.com/science/article/pii/S235271102600169X},
author = {Ronghao Pan and Tomás Bernal-Beltrán and José Antonio García-Díaz and Rafael Valencia-García},
}

Acknowledgments

This work is part of the research project LaTe4PoliticES (PID2022-138099OB-I00), funded by MICIU/AEI/10.13039/501100011033 and the European Regional Development Fund (ERDF/EU - FEDER/UE), “A way of making Europe”.

Mr. Tomás Bernal-Beltrán is supported by the University of Murcia through the predoctoral programme.