Pomoika24/voiceagent-ser-open

VoiceAgent SER — redistributable corpus subset

Dataset Description

This dataset supports speaker-disjoint, multi-corpus speech emotion recognition (SER) research — specifically the open_ser_v1 manifest used to train and evaluate UNIMUSE, a model that fuses a frozen WavLM-Large encoder with hand-crafted acoustic features via cross-attention for six-class emotion classification (neutral, happy, sad, angry, fearful, disgust) across German, Bangla, and Polish speech.

It is not a full copy of every corpus the VoiceAgent project uses for that research. It is the subset whose upstream licenses actually permit redistribution — every other corpus considered (English RAVDESS, MELD, CREMA-D, IEMOCAP, MSP-Podcast, DAIC-WoZ, EmoV-DB) is either restricted to non-commercial/research-only use without redistribution rights, requires a separately signed access agreement, or, in MELD's case, is derived from copyrighted television footage this project has no rights to redistribute. Rather than silently omit that context, this card documents exactly what was left out and why (see "Not included" below), so downstream users know this is a curated legal subset, not the complete training corpus.

Each of the three included corpora keeps its own upstream license and citation requirement — they are bundled together for convenience, not relicensed. Read the per-corpus terms below before use, especially for nemo-polish/, whose CC BY-NC-SA 4.0 license is materially more restrictive (non-commercial, share-alike) than the other two.

Payloads

This repository packages speech-emotion corpora whose data cards provide redistribution-compatible licenses. It has two payloads:

• voiceagent-ser-open.tar — extract it to obtain:
  • emodb/: Berlin EmoDB, German, CC0-1.0 (535 WAV files).
  • subesco/: SUBESCO, Bangla, CC-BY-4.0 (7,000 WAV files).
• nemo-polish/: nEMO, Polish, CC BY-NC-SA 4.0 (4,481 utterances, 2 Parquet shards, audio embedded per row) — mirrored as-is from the upstream amu-cai/nEMO dataset. Non-commercial use only, and any redistribution of this subset specifically must remain under the identical CC BY-NC-SA 4.0 license with attribution to the original authors (see Citation below). This license is materially more restrictive than EmoDB/SUBESCO above — check which subdirectory you're using before relying on its terms.

The original directory layout and metadata tables are preserved in the archive. Emotion labels are corpus-specific and must be normalized before joint training.

Some individual files at the repository root were created by an interrupted bulk upload and are incomplete. Do not use those partial directories; download and extract voiceagent-ser-open.tar instead.

Not included

RAVDESS, EmoV-DB, MELD, CREMA-D, IEMOCAP, MSP-Podcast, DAIC-WoZ, and target-domain recordings are deliberately excluded. Their terms restrict redistribution, require a separate release agreement, or (for MELD) derive from copyrighted third-party media whose rights this project does not hold. Retrieve those datasets from their official providers and consult their terms before use.

Sources and attribution

• Berlin EmoDB: Burkhardt et al. (2005), Interspeech.
• SUBESCO: Islam et al. (2021). Dataset card: https://audeering.github.io/datasets/datasets/subesco.html
• nEMO: Christop, I. (2024). "nEMO: Dataset of Emotional Speech in Polish." Proceedings of LREC-COLING 2024, pp. 12111-12116. https://aclanthology.org/2024.lrec-main.1059. Dataset card: https://huggingface.co/datasets/amu-cai/nEMO. Licensed CC BY-NC-SA 4.0 by its original authors; mirrored here under the same license and terms.

The dataset host must retain the respective upstream notices and licenses.