Recommend BEST-RQ-2.2 and add the project X-ARES comparison table

README.md

@@ -12,6 +12,8 @@ tags:
 ---
 # BEST-RQ-2.1-base
 
+> **Recommended: [BEST-RQ-2.2-base](https://huggingface.co/ltuncay/BEST-RQ-2.2-base)** has the strongest reported X-ARES results in the BEST-RQ-2 family. For new projects, start with that model; see the comparison below.
+
 BEST-RQ-2.1-base is a self-supervised audio encoder trained on AudioSet for
 a configured budget of **200,000 optimizer steps**. It produces **768-dimensional** clip and frame
 embeddings from mono **16 kHz** waveforms, and supports downstream fine-tuning.
@@ -20,6 +22,29 @@ This repository contains the trained encoder, its preprocessing configuration,
 and the custom Transformers implementation. No installation of the research
 repository is needed.
 
+## X-ARES results
+
+Scores on [X-ARES](https://arxiv.org/abs/2505.16369) (0–100, higher is better).
+BEST-RQ (Conformer), BEST-RQ (ViT),
+and all BEST-RQ-2 variants reported below are trained on the **same AudioSet
+split for 200,000 steps**. The pretrained baselines are shown for comparison.
+
+| Model | Speech | Music | Environment | Global Mean | Mean of Means | Training recipe |
+| --- | ---: | ---: | ---: | ---: | ---: | --- |
+| data2vec | 50.62 | 23.24 | 15.41 | 37.83 | 29.76 | Pretrained baseline |
+| wav2vec 2.0 | 41.79 | 34.94 | 29.52 | 37.84 | 35.42 | Pretrained baseline |
+| Whisper | 49.19 | 38.67 | 28.61 | 42.75 | 38.82 | Pretrained baseline |
+| BEST-RQ (Conformer) | 40.43 | 35.58 | 30.81 | 37.43 | 35.60 | Separate codebase (recipe unavailable) |
+| BEST-RQ (ViT) | 32.87 | 41.62 | 34.50 | 34.88 | 36.33 | [best_rq/audioset/default.yaml](https://github.com/LudovicTuncay/audio-embeddings/blob/bb88bf790b1dcf8251c6b38e7a4766534adf33d3/configs/experiment/best_rq/audioset/default.yaml) |
+| BEST-RQ-2 (Interspeech 2026) | 38.49 | 54.40 | 46.39 | 43.21 | 46.43 | [best_rq_2/default.yaml](https://github.com/LudovicTuncay/audio-embeddings/blob/bb88bf790b1dcf8251c6b38e7a4766534adf33d3/configs/experiment/best_rq_2/default.yaml) |
+| BEST-RQ-2.1 | 52.60 | 62.23 | 53.38 | 54.59 | 56.07 | [best_rq_2_1/masking/80.yaml](https://github.com/LudovicTuncay/audio-embeddings/blob/bb88bf790b1dcf8251c6b38e7a4766534adf33d3/configs/experiment/best_rq_2_1/masking/80.yaml) |
+| BEST-RQ-2.2 | **53.78** | **63.90** | **55.71** | **56.11** | **57.80** | [best_rq_2_2/masking/80.yaml](https://github.com/LudovicTuncay/audio-embeddings/blob/bb88bf790b1dcf8251c6b38e7a4766534adf33d3/configs/experiment/best_rq_2_2/masking/80.yaml) |
+
+Global Mean averages all benchmark task scores. Mean of Means gives equal
+weight to the Speech, Music, and Environment category means.
+
+These are the research results reported in the [project README](https://github.com/LudovicTuncay/audio-embeddings/blob/bb88bf790b1dcf8251c6b38e7a4766534adf33d3/README.md), not a new benchmark run of the Transformers exports.
+
 ## Model and training
 
 | Property | Value |

export_manifest.json

@@ -31,7 +31,7 @@
   },
   "files": {
     "CODE_LICENSE": "8fe9e8b749cd4abedabcb3100df445db899e72192394d14cc2dbf24a40811af6",
-    "README.md": "c3ddda7cb48f4d4c8f0f053de4a9786d49818ca1ef8da0a16e0183278934028c",
+    "README.md": "194db8166e421ac7fa8709099f16fa7a9ac476d21d01dd03527aca66ca3a78fc",
     "adapters.py": "ca0f26826763c0e48b7508da732242bb83adfeb4814e389ed3fde93ad648c728",
     "config.json": "88be4a7618663564525588db633cf7461f051e437913187070b854ae8367ac07",
     "configuration_audio.py": "59f3a0b8db0df5af85e677ac33a4431595e9b2eff6d34c1e6a1778dd75a4568d",