Skip to page content
SAVRN
Search Plan a project

Open-weight model · Audio to audio

Mini-BS-RoFormer-V2.5-46.2M

by HiDolen HiDolen/Mini-BS-RoFormer-V2.5-46.2M

Mini-BS-RoFormer-V2.5-46.2M is an open-weight model for audio to audio from HiDolen, released under MIT License. It has 46M parameters. At 16-bit it needs about 0.1 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 127 downloads a month.

Model for the music source separation task, decomposes audio into four stems: bass, drums, other, and vocals.

Parameters46M
Context—
Weights188.9 MB
Licensemit
AccessOpen weights
Monthly Downloads127

Runs On

What it takes to serve Mini-BS-RoFormer-V2.5-46.2M (46M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

PrecisionWeightsMemory neededCheapest setupPer hourAlso fits
16-bit 0.1 GB 0.1 GB 1x MI300X (192 GB)
Vultr
$1.85 1x H100 $1.99 · 1x MI325X $2.00
8-bit 0.0 GB 0.1 GB 1x MI300X (192 GB)
Vultr
$1.85 1x H100 $1.99 · 1x MI325X $2.00
4-bit 0.0 GB 0.0 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 Oct 18, 2026.

Mini-BS-RoFormer-V2.5-46.2M on every accelerator the SAVRN Index prices, at every precision

Model Card

By HiDolen, published under mit, revision 38dafc788346.

Model for the music source separation task, decomposes audio into four stems: bass, drums, other, and vocals. 针对音乐音频分离的 BS-RoFormer 模型,将音频分离为 bass、drums、other、vocals。 在 MUSDB18HQ 数据的 val 集上的性能(单位 SDR,越高越好): 使用的 transformers 库版本为 5.18.0,PyTorch 版本为 2.14.1。为了正常运行模型还需要安装库 soundfile、einops 和 librosa。 GPU 推理: 以上代码会分离出 bass、drums、other 和 vocal 四个轨道。 Mini-BS-RoFormer-V2.5 相比于 V2 的主要改进: 2. 利用注意力池化达成各 stem 共享 mask 解码器 3. 去掉 MUSDB18HQ 数据集中的坏数据进行训练 使用 MUSDB18HQ 数据集的 train 和 test 集进行训练。 学习率 3e-4,以 batchsize=8 训练 246k 步。 对 transformer 层使用 Muon 优化器,其他网络层使用 AdamW 优化器。 - https://github.com/lucidrains/BS-RoFormer - https://arxiv.org/abs/2309.02612 (Music Source Separation with Band-Split RoPE Transformer)

Read HiDolen's full model card

Mini-BS-RoFormer-V2.5

Model for the music source separation task, decomposes audio into four stems: bass, drums, other, and vocals.

针对音乐音频分离的 BS-RoFormer 模型,将音频分离为 bass、drums、other、vocals。

Model Details

  • 模型参数量,46.2M
  • 权重精度,bf16
  • 采样率,44100 Hz
  • 音频分块长度,8 s

在 MUSDB18HQ 数据的 val 集上的性能(单位 SDR,越高越好):

tracks Mini-BS-RoFormer-V2.5-46.2M Mini-BS-RoFormer-V2-46.8M
overall 10.09 10.03
bass 9.71 9.68
drums 10.48 10.58
other 9.19 8.99
vocal 10.97 10.86

Uses

使用的 transformers 库版本为 5.18.0,PyTorch 版本为 2.14.1。为了正常运行模型还需要安装库 soundfile、einops 和 librosa。

GPU 推理:

from transformers import AutoModel
import soundfile
import torch
import librosa

model_name = "HiDolen/Mini-BS-RoFormer-V2.5-46.2M"
model = AutoModel.from_pretrained(
    model_name,
    trust_remote_code=True,
)
model.to("cuda")

# 加载音频
file = "./input.wav"
waveform, sr = librosa.load(file, sr=44100, mono=False)
waveform = torch.tensor(waveform).float()
waveform = waveform.to("cuda")

# 进行推理
result = model.separate(
    waveform,
    batch_size=2,
    verbose=True,
)

# 保存处理结果
for i in range(result.shape[0]):
    soundfile.write(f"separated_stem_{i}.wav", result[i].cpu().numpy().T, 44100)

以上代码会分离出 bass、drums、other 和 vocal 四个轨道。

若想只分离人声和伴奏两轨,在最后保存音频时合并即可:

···

result = model.separate(
    waveform,
    batch_size=2,
    verbose=True,
)

# 合并 bass、drums、other 作为伴奏
instrumental = result[0] + result[1] + result[2]
vocals = result[3]
result = torch.stack([instrumental, vocals], dim=0)
for i in range(result.shape[0]):
    soundfile.write(f"separated_stem_{i}.wav", result[i].cpu().numpy().T, 44100)

Differences from the previous version

Mini-BS-RoFormer-V2.5 相比于 V2 的主要改进:

  1. 损失改为频域听阈加权 L1
  2. 利用注意力池化达成各 stem 共享 mask 解码器
  3. 去掉 MUSDB18HQ 数据集中的坏数据进行训练
  4. 使用更宽更浅的模型超参配置

Training Details

使用 MUSDB18HQ 数据集的 train 和 test 集进行训练。

学习率 3e-4,以 batch_size=8 训练 246k 步。

对 transformer 层使用 Muon 优化器,其他网络层使用 AdamW 优化器。

Acknowledgments

  • https://github.com/lucidrains/BS-RoFormer
  • https://arxiv.org/abs/2309.02612 (Music Source Separation with Band-Split RoPE Transformer)

Configuration

Architecture
BSRoformerForMaskedEstimation
Layers
6
Hidden size
512
Feed-forward size
1,024
Attention heads
8
Key/value heads
4
Head dimension
64
Model type
bs_roformer

Identity and Version

Repository
HiDolen/Mini-BS-RoFormer-V2.5-46.2M
Publisher
HiDolen
Task
Audio to audio
Modality
Audio
Library
transformers
Parameters
46M parameters
Languages
Not stated by the source
Revision
38dafc78834607aec57d8c0cb3568523f9ea5563
First published
2026-10-10
Last updated
2026-10-17

Files and Weights

9 files, 189.1 MB in total. The weights are 2 files totalling 188.9 MB in onnx, safetensors.

Weights2 files · 188.9 MB
Configuration4 files · 49.2 KB
Documentation1 file · 3.2 KB
Other1 file · 107.4 KB
Repository1 file · 1.5 KB
Every file
FileTypeSizeSHA-256
model.safetensorsWeights92.4 MB 0f173c42b871
onnx/freq_weights.onnxWeights96.5 MB c019650a7fb9
config.jsonConfiguration2.1 KB —
configuration_bs_roformer.pyConfiguration3.5 KB —
modeling_bs_roformer.pyConfiguration40.8 KB —
onnx/onnx_inference_example.pyConfiguration2.7 KB —
README.mdDocumentation3.2 KB —
training_process.ipynbOther107.4 KB —
.gitattributesRepository1.5 KB —

License and Download

License
mit
Access
Open weights, no gate
Download size
188.9 MB
Download from HiDolen

Released by HiDolen through its official repository on Hugging Face. Read the license.

Built From

  • Described by arXiv:2309.02612
  • Trained on (disclosed) CLAPv2/MUSDB18-HQ

Memory Requirements

PrecisionWeights in memory
As published188.9 MB
16-bit0.1 GB
8-bit0.0 GB
4-bit0.0 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About Mini-BS-RoFormer-V2.5-46.2M

How much GPU memory does Mini-BS-RoFormer-V2.5-46.2M need?

About 0.1 GB at 16-bit and 0 GB at 4-bit: the weights (46M parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run Mini-BS-RoFormer-V2.5-46.2M on?

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.

Can I use Mini-BS-RoFormer-V2.5-46.2M commercially?

Yes. Mini-BS-RoFormer-V2.5-46.2M is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

Similar Models

Model · Audio to audio

bigvgan_v2_22khz_80band_256x

NVIDIA

[[Paper]](https://arxiv.org/abs/2206.04658) - [[Code]](https://github.com/NVIDIA/BigVGAN) - [[Showcase]](https://bigvgan-demo.github.io/) - [[Project Page]](https://research.nvidia.com/labs/adlr/projects/bigvgan/) - [[Weights]](https://huggingface.co/collections/nvidia/bigvgan-66959df3d97fd7d98d97dc9a) - [[Demo]](https://huggingface.co/spaces/nvidia/BigVGAN) - General refactor and code improvements for improved readability. - Fully fused CUDA kernel of anti-alised activation (upsampling + activation + downsampling) with inference speed benchmark. - We provide pretrained checkpoints of BigVGAN-v2 using diverse audio configurations, supporting up to 44 kHz sampling rate and 512x upsampling…

Open weights mit PyTorch

Model · Audio to audio

bigvgan_v2_44khz_128band_512x

NVIDIA

[[Paper]](https://arxiv.org/abs/2206.04658) - [[Code]](https://github.com/NVIDIA/BigVGAN) - [[Showcase]](https://bigvgan-demo.github.io/) - [[Project Page]](https://research.nvidia.com/labs/adlr/projects/bigvgan/) - [[Weights]](https://huggingface.co/collections/nvidia/bigvgan-66959df3d97fd7d98d97dc9a) - [[Demo]](https://huggingface.co/spaces/nvidia/BigVGAN) - General refactor and code improvements for improved readability. - Fully fused CUDA kernel of anti-alised activation (upsampling + activation + downsampling) with inference speed benchmark. - We provide pretrained checkpoints of BigVGAN-v2 using diverse audio configurations, supporting up to 44 kHz sampling rate and 512x upsampling…

Open weights mit PyTorch

Descript Audio Codec running on-device on the LiteRT CompiledModel GPU (ML Drift). The convolutional encoder/decoder run on the GPU; the RVQ runs on CPU. 43:1 compression (1 s → 12×50 codes), RTF ≈ 0.82 (faster than real-time) on Pixel 8a. - dac16khzencoderfp16.tflite (43 MB) — audio[1,1,16000] → latent[1,1024,50], GPU. - dac16khzdeconlyzsfp16.tflite (105 MB) — latent[1,1024,50] → audio, GPU. - dacrvq.bin (1.2 MB) — RVQ weights (12 codebooks) for the CPU quantizer (float32 LE). encoder 367/367 + decoder 398/398 nodes on the LiteRT GPU delegate (LITERTCL, 1 partition, no CPU fallback); warm RTF ~0.82; reconstruction corr 1.0 vs PyTorch DAC. The decoder's ConvTranspose1d are rewritten to a…

Open weights mit litert

Model · Audio to audio

ArkEcho-RVC-M3-C-F

Ethernos

本仓库(Repository)所包含的所有人工神经网络(Artificial Neural Network)权重文件(.pth/.index)、训练日志及相关代码,均为计算声学(Computational Acoustics)与深度学习(Deep Learning)领域的技术研究实验产物。 这些文件本质上是高维张量(High-dimensional Tensors)的数值序列,通过随机梯度下降(SGD)与反向传播算法(Backpropagation)对公开可获取的音频数据进行统计建模(Statistical Modeling)得到。其技术形态与数字图像处理中的卷积核(Convolution Kernels)、自然语言处理中的词向量(Word Embeddings)并无本质差异。 本仓库不构成对任何第三方知识产权的故意侵犯,所有代码遵循 MIT License 开源协议,模型权重文件仅作为技术实现的副产品(By-products)存在。 本技术实验所使用的训练数据集(Training Dataset)包含以下角色的语音样本: - Mon3tr(モンスター)、Mon3tr 中文语音 等 - 声源版权归属:上海鹰角网络科技有限公司(Hypergryph Network Technology Co., Ltd.)及其关联公司 - 原始作品:《明日方舟》(Arknights) 上述角色的声音版权、肖像权、姓名权及相关知识产权均完全归属于鹰角网络及其合法授权方。本实验仅基于已公开发布的游戏内语音资源进行技术层面的信号处理(Signal Processing)与特征提取(Feature…

Open weights cc-by-nc-4.0

Model · Audio to audio

MOSS-v2-12to32-Enhancer-60M

LAION eV

A small causal Qwen3 model that predicts 32 MOSS v2 audio-codebook indices from the 12 indices of the current frame and up to ten previously generated HQ frames. It is trained from scratch on German, English, Spanish and French speech. This is an experimental enhancement model. Validation cross entropy measures token prediction under teacher forcing. It does not establish an improvement in perceived audio quality. Use the generated-history listening evaluation to judge speech content, speaker identity, artifacts and long-utterance stability. Root inference weights: step 64,699, selected by held-out extra-20-codebook cross entropy 5.818123. - Step 12,940: 100 originals / 200 input cases…

Open weights cc-by-nc-4.0 pytorch

Model · Audio to audio

gtcrn-coreai

Max Farrell

This release mirrors maxffarrell/gtcrn-coreai at source commit 5a5b0529f89c7146fc196f602ed2635adc681374. Choose DNS3 for general speech enhancement; VCTK-DEMAND is provided for matching that training corpus and reproducing related experiments. Download the complete repository, preserving the.aimodel bundle directories: For deployment you can download only exports/, LICENSE, and THIRDPARTYNOTICES.md using hf download --include. The numerical and latency results below were measured in August 2026 on the specified beta software stack; they have not been remeasured for newer releases. A reproducible, streaming Core AI conversion of GTCRN, the 48.2K-parameter speech-enhancement network…

Open weights mit