Update README.md

README.md

@@ -31,50 +31,74 @@ tags:
 - zero-shot
 - unsupervised-clustering-generalization
 ---
-# VocSim: Zero-Shot Audio Similarity Benchmark
+# VocSim: A Benchmark for Zero-Shot Audio Similarity
 
-## Dataset Description
+[![License: CC BY 4.0](https://img.shields.io/badge/License-CC%20BY%204.0-blue.svg)](https://creativecommons.org/licenses/by/4.0/)
+[![Dataset Size](https://img.shields.io/badge/Dataset%20Size-5.45%20GB-orange)](https://huggingface.co/datasets/anonymous-submission000/vocsim)
+[![Number of Examples](https://img.shields.io/badge/Examples-114k-brightgreen)](https://huggingface.co/datasets/anonymous-submission000/vocsim)
 
-**Repository:** [LINK UPON DOI]
-**Paper:** [VocSim: Zero-Shot Audio Similarity Benchmark for Neural Embeddings (LINK UPON DOI)]()
-**Point of Contact:** Anonymous at the moment
+**VocSim** is a large-scale benchmark dataset meticulously crafted to evaluate the generalization capabilities of neural audio embeddings for **zero-shot audio similarity tasks**. It challenges models to recognize fine-grained acoustic similarity between sounds they haven't been explicitly trained to classify together.
 
-### Overview
+**Repository:** [GitHub Link - Add Upon DOI]
+**Paper:** [VocSim: Zero-Shot Audio Similarity Benchmark for Neural Embeddings - Link Upon DOI]()
+**Point of Contact:** Anonymous Authors (initially)
 
-**VocSim** is a large-scale benchmark dataset designed to evaluate the generalization capabilities of neural audio embeddings for **zero-shot audio similarity**. 
+---
+
+## 🎯 Dataset Goal: Evaluating Generalization
+
+The core purpose of VocSim is to provide a standardized testbed for assessing how well learned audio representations can capture the notion of "sameness" between sounds based purely on their acoustic properties, without relying on supervised training signals for the specific similarity task being evaluated.
+
+## 🧩 Dataset Structure: Diverse Acoustic Challenges
+
+VocSim achieves its goal by aggregating **19 distinct subsets** (15 publicly available, 4 reserved as blind test sets) that introduce significant variability across key acoustic and structural dimensions:
 
-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:** Spans human speech (phonemes, words, utterances), intricate birdsong (calls, syllables from multiple species), otter calls, and common environmental events.
+*   ⏱️ **Duration:** Ranges from fleeting sub-100ms syllables to multi-second recordings.
+*   📊 **Class Structure:** Includes subsets with a few well-populated classes (e.g., environmental sounds) alongside subsets with thousands of classes having only a handful of examples each (e.g., rare words or syllables).
+*   🔊 **Acoustic Conditions:** Features both clean, studio-quality recordings (like TIMIT portions) and challenging, noisy, naturalistic recordings (e.g., AMI meetings, field recordings of animals).
+*   🧬 **Intra-class Variation:** Naturally incorporates real-world differences in loudness, speaking/singing rate, speaker identities, animal individuals, and recording environments within the same semantic `label`.
 
-*   **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.
+This diversity forces embedding models to learn robust representations that are invariant to nuisance factors while remaining sensitive to the core acoustic characteristics defining a sound class.
 
-### Supported Tasks and Leaderboards
+### Key Metrics
 
-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 metrics sensitive to retrieval quality and cluster coherence:
 
-Performance is measured using:
+*   **RA@k (Retrieval Accuracy @k):** The fraction of the *k* retrieved nearest neighbors that share the same `label` as the query sample. *Higher is better.* (e.g., RA@1 checks if the single closest neighbor is correct).
+*   **CSCF (Cluster Separation Confusion Fraction):** Estimates the proportion of samples that are, on average, closer to embeddings from *other* classes than to embeddings from their *own* class. Reflects how well-separated the class clusters are in the embedding space. *Lower is better.*
 
-*   **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.
+(See Section 5 of the paper for detailed definitions and analysis.)
 
-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).
+### Other Potential Uses
 
-[Leaderboard upon DOI]
+Beyond zero-shot similarity, VocSim embeddings can be evaluated on:
+
+*   Unsupervised Clustering (e.g., using metrics like NMI or ARI).
+*   As input features for downstream supervised classification tasks.
+
+## 🏆 Leaderboard
+
+Track the performance of different embedding models on VocSim:
+
+➡️ **[VocSim Leaderboard on Papers With Code](https://paperswithcode.com/dataset/audiosim)** ⬅️
+---
 
+## 💾 Data Format
 
 ### Data Instances
 
-A typical data point consists of an audio sample and its associated metadata:
+Each data point in the dataset includes the audio waveform and associated metadata. All audio is standardized to **16kHz mono**.
 
 ```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
-}
+  'audio': {
+      'path': None, # Path may be None if loaded directly
+      'array': array([-0.00024, -0.00048, ..., 0.00012], dtype=float32), # Waveform
+      'sampling_rate': 16000 # Always 16kHz
+    },
+  'subset': 'HW1',             # Origin subset ID (e.g., Human Words 1)
+  'speaker': 'speaker_id_abc', # Identifier for speaker, animal, or session. 'N/A' if not applicable.
+  'label': 'hello'              # Ground truth class label (word, phone, syllable type, etc.)
+}
+```