LAION's Emolia Balanced 5M Subset is a curated collection of approximately 5.26 million speech samples, filtered from the larger 80.5 million-sample Emolia dataset. The subset was created by selecting samples that meet specific emotion score thresholds based on the Emonet taxonomy of 40 emotions. It is packaged as WebDataset-compatible tar shards for direct use in training pipelines and was last updated in April 2026.
Use Cases
- Train speech emotion recognition models based on the 40 emotion annotation scores.
- Benchmark model performance on a balanced subset of emotional speech data.
- Fine-tune audio foundation models for affective computing tasks.
- Conduct research on the distribution and co-occurrence of emotions in speech.
Strengths
- Contains approximately 5.26 million samples, providing substantial scale for training.
- Explicitly balanced based on emotion score thresholds, which may help mitigate class imbalance.
- Packaged in WebDataset-compatible format for efficient integration into training pipelines.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is an approximate figure; exact sample composition requires verification.
- Data may reflect bias inherent to the source collection methods and original annotation process.
Provenance
- Source
- LAION, derived from the larger laion/Emolia dataset.
- Collection Method
- Samples were filtered from the original 80.5M collection based on meeting emotion score thresholds.
- Time Range
- null
- Freshness
- Last updated 2026-04-19 23:13:56; freshness should be verified.
- Geography
- null