Update README.md

README.md

@@ -34,4 +34,51 @@ tags:
 pretty_name: VocSim
 size_categories:
 - 100K<n<1M
----
+---
+# VocSim: Zero-Shot Audio Similarity Benchmark
+
+## Dataset Description
+
+**Repository:** [LINK UPON DOI]
+**Paper:** [VocSim: Zero-Shot Audio Similarity Benchmark for Neural Embeddings (LINK UPON DOI)]()
+**Point of Contact:** Anonymous at the moment
+
+### Overview
+
+**VocSim** is a large-scale benchmark dataset designed to evaluate the generalization capabilities of neural audio embeddings for **zero-shot audio similarity**. 
+
+The dataset is intentionally structured into distinct subsets (19 in total, 15 available here and 4 blind test set) to represent variability in:
+
+*   **Sound Source:** Human speech (phonemes, words, utterances), birdsong (calls, syllables from multiple species), otter calls, environmental events.
+*   **Duration:** From <100ms syllables to multi-second utterances.
+*   **Class Structure:** Subsets range from few, well-populated classes to thousands of classes with few examples.
+*   **Acoustic Conditions:** Clean recordings (e.g., TIMIT) contrasted with noisy, naturalistic recordings (e.g., AMI meetings, field recordings).
+*   **Intra-class Variation:** Natural differences in loudness, rate, speaker/animal identity within classes.
+
+### Supported Tasks and Leaderboards
+
+The primary intended task is **zero-shot audio similarity evaluation**. Given an audio sample, the goal is to retrieve other samples with the same `label` (representing the same underlying sound class, e.g., word, syllable type) based solely on comparing their embeddings using a distance metric (e.g., Cosine distance).
+
+Performance is measured using:
+
+*   **RA@k (Retrieval Accuracy @k):** Measures the fraction of the k-nearest neighbors that belong to the same class as the query sample. Higher is better.
+*   **CSCF (Cluster Separation Confusion Fraction):** Measures the fraction of samples that are, on average, closer to samples from *other* classes than to samples from their *own* class. Lower is better.
+
+The dataset can also be used for evaluating embeddings in downstream tasks like unsupervised clustering or as features for supervised classification (see Section 5 of the paper).
+
+[Leaderboard upon DOI]
+
+
+### Data Instances
+
+A typical data point consists of an audio sample and its associated metadata:
+
+```python
+{
+  'audio': {'path': '/path/to/audio.wav', 'array': array([-0.00024414, -0.00048828, ...,  0.00012207], dtype=float32), 'sampling_rate': 16000},
+  'subset': 'HW1', # Example: Human Words from TIMIT
+  'speaker': 'speaker_id_abc', # Or animal ID, or 'N/A'
+  'label': 'hello', # Example label (could be a word, phone, class ID)
+  'index': 12345,
+  'sampling_rate': 16000
+}