🤫😀 NVIDIA NemotronLabs RE-USE-laughter: Multilingual Universal Speech Enhancement

Model Overview

Description:

NVIDIA NemotronLabs RE-USE-laughter is a 9.6-million-parameter multilingual universal speech enhancement (USE) model designed to restore speech quality across a wide range of real-world degradations while preserving speech fidelity. It has the following desirable properties:

• This model better preserves laughter while restoring the quality of the input speech.
• Robustness to diverse degradations, including additive noise, reverberation, clipping, bandwidth limitation, codec artifacts, packet loss and low-quality mics.
• Support for multiple input sampling rates, including 8, 16, 22.05, 24, 32, 44.1, and 48 kHz.
• Strong language-agnostic capability, enabling effective performance across different languages.

(Note: For overlapping speech, this model is trained to preserve the main speaker while suppressing other speakers.)

NVIDIA NemotronLabs RE-USE-laughter was developed by NVIDIA as a part of RE-USE.

This model is ready for commercial use.

Performance

Comparison with other open-source USE models on the (subsets) of the URGENT 2025 non-blind test set, referenced against original anechoic clean speech.

SBERT: SpeechBERTScore, LPS: Levenshtein Phoneme Similarity,

Spk: Speaker similarity (ECAPA-TDNN), CAcc: Character accuracy of ASR (OWSM v3.1)

See the performance comparison between NemotronLabs RE-USE and other models, including Sidon, UniPASE, UniSE, Resemble Enhance, and ClearVoice here

Usage

Directly try our Gradio Interactive Demo by uploading your noisy audio/video !!

Environment Setup

1. (For Mamba setup)Pre-built Docker environments can be downloaded here to simplify Mamba setup.

2. If you need bandwidth extension:

pip install resampy 

3. Download and navigate to the HuggingFace repository:

huggingface-cli download nvidia/NVIDIA-NemotronLabs-RE-USE-laughter  --local-dir ./NemotronLabs_REUSE_laughter --local-dir-use-symlinks False
cd ./NemotronLabs_REUSE_laughter

Inference

Follow the simple steps below to generate enhanced speech using our model:

1. Place your noisy speech files in the folder noisy_audio/
2. Run the following command:

sh inference.sh

3. The enhanced speech files will be saved in enhanced_audio/.

That's all !

Note:

a. You can enable bandwidth extension by setting the target bandwidth using the BWE argument in the script.

---

If your noisy speech files are long (e.g., >30s, compared to our training data, which consists of clips < 10s) or cause GPU out-of-memory (OOM) errors, please use the following procedure instead:

1. Place your long noisy speech files in the folder long_noisy_audio/
2. Run the following command:

sh inference_chunk.sh

3. The enhanced speech files will be saved in Long_enhanced_audio/.

Note:

a. You can also configure the chunk_size_in_seconds and hop_length_portion directly in the script.

License/Terms of Use:

OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)

This project is currently not accepting contributions.

Deployment Geography:

Global

Use Case:

Researchers and general users can use this model to enhance the quality of speech recordings across a wide range of applications. For example, content creators can use it to improve the speech quality of their recordings, while researchers can use it to enhance speech datasets for training speech generation models, such as text-to-speech (TTS) and speech large language models (Speech LLMs).

Release Date:

Hugging Face 10/22/2026

Reference(s):

Rethinking Training Targets, Architectures and Data Quality for Universal Speech Enhancement (NeurIPS 2026)
One Model, Many Latencies: Universal Speech Enhancement for Diverse Real-Time Applications (IEEE SLT 2026)

Model Architecture:

Architecture Type: Convolutional encoder, Convolutional decoder, and Mamba for time–frequency modeling
Network Architecture: Bi-directional Mamba with 30 layers
Number of model parameters: 9.6M (9.6*10^6)

Input:

Input Type(s): Audio
Input Format(s): .wav files
Input Parameters: One-Dimensional (1D)
Other Properties Related to Input: 8000 Hz - 48000 Hz Mono-channel Audio

Output:

Output Type(s): Audio
Output Format: .wav files
Output Parameters: One-Dimensional (1D)
Other Properties Related to Output: 8000 Hz - 48000 Hz Mono-channel Audio

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration:

Runtime Engine(s): N/A
Supported Hardware Microarchitecture Compatibility: NVIDIA Ampere
Supported Operating System(s): Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Model Version(s):

NVIDIA-NemotronLabs-RE-USE-laughter_v1

Training, Testing, and Evaluation Datasets:

Training Dataset:

Data Modality: Audio

Audio Training Data Size: ~2500 hours
Data Collection Method by dataset: Synthetic
Labeling Method by dataset: Synthetic
Properties (Quantity, Dataset Descriptions, Sensor(s)): The size of the speech data is roughly 2500 hours. Single channel, single speaker, 8kHz - 48kHz sampled data.

• LibriVox data from DNS5 challenge (EN) (~350 hours of speech data)
• LibriTTS (EN) (~200 hours of speech data)
• VCTK (EN) (~80 hours of speech data)
• WSJ (EN) (~85 hours of speech data)
• Multilingual Librispeech (De, En, Es, Fr) (~450 hours of speech data)
• CommonVoice 19.0 (De, En, Es, Fr, zh-CN) (~1300 hours of speech data)
• Riva Speakers: Internal Dataset
• Audioset+FreeSound noise in DNS5 challenge (~180 hours of noise data)
• Simulated RIRs from DNS5 challenge (~60k samples of room impulse response)
• MicIRP (~70 samples of microphone impulse response)
• VocalSound

Testing/Evaluation Dataset:

Data Collection Method by dataset: Synthetic
Labeling Method by dataset: Synthetic
Properties (Quantity, Dataset Descriptions, Sensor(s)): URGENT 2025 non-blind test set, evaluated against the original clean speech as the reference, without convolution with early reflections.

Inference:

Acceleration Engine: Triton

Hardware Requirements (GPU Architecture, Model): NVIDIA Ampere - A100

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Citation

Please consider to cite our paper and this framework, if they are helpful in your research.

@article{fu2026rethinking,
  title={Rethinking Training Targets, Architectures and Data Quality for Universal Speech Enhancement},
  author={Fu, Szu-Wei and Chao, Rong and Yang, Xuesong and Huang, Sung-Feng and Zezario, Ryandhimas E and Nasretdinov, Rauf and Juki{\'c}, Ante and Tsao, Yu and Wang, Yu-Chiang Frank},
  journal={arXiv preprint arXiv:2603.02641},
  year={2026}
}

@article{fu2026one,
  title={One Model, Many Latencies: Universal Speech Enhancement for Diverse Real-Time Applications},
  author={Fu, Szu-Wei and Chao, Rong and Yang, Xuesong and Huang, Sung-Feng and Juki{\'c}, Ante and Tsao, Yu and Wang, Yu-Chiang Frank},
  journal={arXiv preprint arXiv:2606.25621},
  year={2026}
}