mamtra_inference

KO | EN

gitlite — search

#javascript #python #hacktoberfest #typescript #llm #machine-learning #rust #deep-learning #docker #java #cli #kubernetes #pytorch #ios

mamtra_inference

(★ 10)

[Interspeech 2026] The official inference code of MamTra

mamtra_inference Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# MamTra: A Hybrid Mamba-Transformer Backbone for Speech Synthesis

</div>

---

## Authors

| [Tan Dat Nguyen](https://signofthefour.github.io/)<sup>1</sup> | [Sangmin Bae](https://www.raymin0223.com/)<sup>1</sup> | [Joon Son Chung](https://mm.kaist.ac.kr/joon/)<sup>1</sup> | [Ji-Hoon Kim](https://mintlab.cau.ac.kr/members/faculty)<sup>2†</sup> |
|:---:|:---:|:---:|:---:|

<sup>1</sup> Korea Advanced Institute of Science and Technology (KAIST), South Korea  
<sup>2</sup> Chung-Ang University, South Korea  
<sup>†</sup> Corresponding author &nbsp;·&nbsp; [tandat.kaist@kaist.ac.kr](mailto:tandat.kaist@kaist.ac.kr)

</div>

---

## Overview

MamTra introduces a hybrid Mamba-Transformer backbone for text-to-speech synthesis, combining the linear-time sequence modeling of Mamba with the global context modeling of Transformer attention. The result is a system that achieves strong synthesis quality while remaining efficient and hardware-friendly.

> **Note:** MamTra depends on Mamba and associated CUDA extensions, which may require careful environment setup — particularly on machines with non-standard CUDA configurations. If you encounter installation issues, please open a GitHub issue and we will do our best to help.

---

## 🚀 Quickstart

### 1. Environment Setup

**Python 3.10** is required due to a dependency on CosyVoice's `ttsfrd`. We recommend isolating the environment to avoid package conflicts:

```bash
conda create -n mamtra python=3.10 -y
conda activate mamtra
```

**CUDA 11.6 or later** is required to build the Mamba CUDA kernels. If your system CUDA version is too old, you can install a compatible version via conda:

```bash
conda install nvidia/label/cuda-11.8.0::cuda-nvcc
conda install nvidia/label/cuda-11.8.0::cuda
```

Install CuDNN alongside:

```bash
conda install anaconda::cudnn
# alternative if the above fails:
conda install -c conda-forge cudnn
```

When using the conda-managed CUDA, make sure to set `LD_LIBRARY_PATH` accordingly — see [`scripts/uv_inference_hybrid_11.sh`](scripts/uv_inference_hybrid_11.sh) for an example.

### 2. Install Dependencies

We recommend [uv](https://github.com/astral-sh/uv) for dependency management. It handles the version conflicts that commonly arise when installing Mamba, `causal_conv1d`, and CosyVoice together:

```bash
pip install -U uv
uv sync
uv sync --extra compile   # required for the hybrid model
```

---

## Inference

### 0. Download Checkpoint

The MamTra 1:1 checkpoint is available on HuggingFace. Download it with:

```bash
huggingface-cli download blueskyheaven/mamtra11 \
  --local-dir huggingface_model_dir \
  --local-dir-use-symlinks False
```

> Additional model variants will be released in the near future.

Once downloaded, set the path as the `MODEL_DIR` variable in the inference script of your choice.

### 1. Single Sample Inference

```bash
bash scripts/single_inference.sh
```

### 2. Batch Inference

```bash
bash scripts/uv_inference_hybrid_11.sh
```

---

## Citation

If you find MamTra useful in your research, please consider citing:

```bibtex
@article{nguyen2025mamtra,
  title   = {MamTra: A Hybrid Mamba-Transformer Backbone for Speech Synthesis},
  author  = {Nguyen, Tan Dat and Bae, Sangmin and Chung, Joon Son and Kim, Ji-Hoon},
  journal = {arXiv preprint arXiv:2603.12342},
  year    = {2025}
}
```

mamtra_inference

mamtra_inference Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

mamtra_inference

mamtra_inference Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README