Dataset card written from measured statistics

README.md

@@ -1,16 +1,14 @@
 ---
+license: mit
+language:
+- ru
 annotations_creators:
 - crowdsourced
 language_creators:
 - expert-generated
-- crowdsourced
-language:
-- ru
-license:
-- mit
 multilinguality:
 - monolingual
-pretty_name: Russian Emotional Phonetic Voices Small
+pretty_name: Russian Emotional Phonetic Voices — Small
 size_categories:
 - 1K<n<10K
 source_datasets:
@@ -19,10 +17,62 @@ task_categories:
 - audio-classification
 task_ids:
 - audio-emotion-recognition
+tags:
+- emotion-recognition
+- russian
+- speech
 ---
 
-# Citations
+![Russian Emotional Phonetic Voices — Small](assets/banner.svg)
+
+# Russian Emotional Phonetic Voices — Small
+
+The compact REPV subset.
+
+## How it was collected
+
+REPV was gathered by crowdsourcing rather than in a studio: around **200 different speakers** for the full set and about **50** for REPV-S. Recording conditions therefore vary from contributor to contributor, which makes it harder than RESD and closer to what a microphone in the wild actually receives.
+
+## Splits
+
+| Split | Rows | Hours | Mean clip |
+|---|---:|---:|---:|
+| `train` | 112 | 0.12 | 3.7 s |
+| `test` | 28 | 0.03 | 3.9 s |
+
+<img src="assets/classes.svg" alt="Class distribution" width="760">
+
+## Fields
+
+| Column | Meaning |
+|---|---|
+| `path` | Original file path |
+| `file` | Source file name |
+| `gender` | Speaker gender as reported by the contributor |
+| `emotion` | Emotion label of the recording |
+| `speech` | Audio |
+
+Gender is close to even in `train`: 58 f, 54 m.
+
+> [!NOTE]
+> The label set is **not** the seven-class one used by RESD and the Aniemore models. REPV has five: `anger`, `enthusiasm`, `happiness`, `sadness` and `tiredness`. `tiredness` appears nowhere else in the library, and `neutral`, `fear` and `disgust` are absent here.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("Aniemore/REPV-S")
+print(ds["train"][0]["emotion"])
 ```
+
+## Limitations
+
+Crowdsourced audio varies in microphone, room and level, and the whole set is 0.1 hours — small enough that a single split can move a score by several points. `REPV-S` in particular holds 140 clips in total and is meant for smoke tests rather than for measuring anything.
+
+## Citation
+
+```bibtex
 @misc{Aniemore,
   author = {Артем Аментес, Илья Лубенец, Никита Давидчук},
   title = {Открытая библиотека искусственного интеллекта для анализа и выявления эмоциональных оттенков речи человека},
@@ -32,4 +82,8 @@ task_ids:
   howpublished = {\url{https://huggingface.com/aniemore/Aniemore}},
   email = {hello@socialcode.ru}
 }
-```
+```
+
+## License
+
+MIT.