Search public pages, research tools, and SAVRN solutions.
This is the official repository for the paper "MuQ: Self-Supervised Music Representation Learning with Mel Residual Vector Quantization". For more detailed information, we strongly recommend referring to https://github.com/tencent-ailab/MuQ and the paper).
What it takes to serve MuQ-large-msd-iter (333M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 0.7 GB | 0.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.3 GB | 0.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.2 GB | 0.2 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
Music, not speech, is the signal here: a 333M parameter representation model pre-trained by self-supervised learning with Mel residual vector quantization, the method in arXiv:2501.01108. Memory barely registers: 0.8 GB at 16-bit, 0.4 GB at 8-bit, 0.2 GB at 4-bit. Our cheapest listed card, one MI300X with 192 GB at $1.85 per hour on-demand, could hold hundreds of copies, so put it on a card already doing other audio work.
The license is where a buyer stops. CC BY-NC 4.0 permits sharing and adapting with credit for non-commercial purposes only; a research group is fine, but anything that earns revenue needs separate permission from the rights holder first. The configuration declares no library or architecture, so loading follows the publisher's code in the tencent-ailab MuQ repository, and the 2.67 GB of weights were last updated August 15, 2025, eight months after the December 17, 2024 release.
This is the official repository for the paper "MuQ: Self-Supervised Music Representation Learning with Mel Residual Vector Quantization". For more detailed information, we strongly recommend referring to https://github.com/tencent-ailab/MuQ and the paper). In this repo, the following models are released: - MuQ(see this link): A large music foundation model pre-trained via Self-Supervised Learning (SSL), achieving SOTA in various MIR tasks. - MuQ-MuLan(see this link): A music-text joint embedding model trained via contrastive learning, supporting both English and Chinese texts. To begin with, please use pip to install the official muq lib, and ensure that your python>=3.8: To extract music…
Excerpt from the card by MuQ, licensed cc-by-nc-4.0.
5 files, 2.7 GB in total. The weights are 2 files totalling 2.7 GB in bin, safetensors.
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 1.3 GB | 273febab2be0 |
| pytorch_model.bin | Weights | 1.3 GB | 334df3de2832 |
| config.json | Configuration | 3.1 KB | — |
| README.md | Documentation | 4.4 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
Released by MuQ through its official repository on Hugging Face. Read the license.
| Precision | Weights in memory |
|---|---|
| As published | 2.7 GB |
| 16-bit | 0.7 GB |
| 8-bit | 0.3 GB |
| 4-bit | 0.2 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
About 0.8 GB at 16-bit and 0.2 GB at 4-bit: the weights (333M parameters) plus a working margin. A long context needs more.
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Not without separate permission. MuQ-large-msd-iter is released under Creative Commons Attribution-NonCommercial 4.0. CC BY-NC 4.0 permits sharing and adapting with credit for non-commercial purposes only. Commercial use needs separate permission from the rights holder.
The model was trained on MSP-Podcast for the Odyssey 2024 Emotion Recognition competition baseline This particular model is the multi-attributed based model which predict arousal, dominance and valence in a range of approximately 0...1. CCC based on Test3 and Development sets of the Odyssey Competition
The model expects a raw audio signal as input and outputs predictions for age in a range of approximately 0...1 (0...100 years) and gender expressing the probababilty for being child, female, or male. In addition, it also provides the pooled states of the last transformer layer. The model was created by fine-tuning Wav2Vec2-Large-Robust Timit and For this version of the model we trained all 24 transformer layers. An ONNX export of the model is available from Further details are given in the associated paper and tutorial.
Speech emotion recognition for Russian over seven classes: anger, disgust, enthusiasm, fear, happiness, neutral, sadness. Fine-tuned from jonatasgrosman/expw2v2truwavlms363 on Aniemore/resd. Audio resampled to 16 kHz mono, clips capped at 12 s, normalized per utterance, padding masked. UA is macro-averaged recall, WA is accuracy, F1 is macro-averaged. All three test sets went through the same harness, so the rows are comparable to each other. The RESD split matches fold 1 of EmoBox bit for bit. The top entry there is WavLM-large at WA 56.47 / UA 55.87 / F1 55.82. These numbers are higher, but the training protocol differs — EmoBox freezes the encoder and trains a probe, this is a full…
Model · Audio classification
The model is a fine-tuned version of jonatasgrosman/wav2vec2-large-xlsr-53-english for a Speech Emotion Recognition (SER) task. The dataset used to fine-tune the original pre-trained model is the RAVDESS dataset. This dataset provides 1440 samples of recordings from actors performing on 8 different emotions in English, which are: It achieves the following results on the evaluation set: The following hyperparameters were used during training: - learningrate: 0.0001 - trainbatchsize: 4 - evalbatchsize: 4 - gradientaccumulationsteps: 2 - totaltrainbatchsize: 8 - lrschedulertype: linear - numepochs: 3 - mixedprecisiontraining: Native AMP Any doubt, contact me on Twitter. - Transformers 4.8.2…
Speech emotion recognition for Russian over seven classes: anger, disgust, enthusiasm, fear, happiness, neutral, sadness. Fine-tuned from jonatasgrosman/wav2vec2-large-xlsr-53-russian on Aniemore/resd. Audio resampled to 16 kHz mono, clips capped at 12 s, normalized per utterance, padding masked. UA is macro-averaged recall, WA is accuracy, F1 is macro-averaged. All three test sets went through the same harness, so the rows are comparable to each other. The RESD split matches fold 1 of EmoBox bit for bit. The top entry there is WavLM-large at WA 56.47 / UA 55.87 / F1 55.82. These numbers are higher, but the training protocol differs — EmoBox freezes the encoder and trains a probe, this is a…
The pre-trained model is this one - facebook/hubert-large-ls960-ft The DUSHA dataset used can be found here Fine-tuned in Google Colab using Pro account with A100 GPU Freezed all layers exept projector, classifier and all 24 HubertEncoderLayerStableLayerNorm layers Used half of the train dataset - 2 epochs - train batch size = 8 - eval batch size = 8 - gradient accumulation steps = 4 - learning rate = 5e-5 without warm up and decay Achieved - accuracy = 0.86 - balanced = 0.76 - macro f1 score = 0.81 on test set, improving accucary and f1 score compared to dataset baseline