How to Run NVIDIA RE-USE Speech Enhancement on a CPU

How do you run NVIDIA RE-USE speech enhancement on a CPU?

Install the optional resampy package, download the model repository from Hugging Face with huggingface-cli, change into the folder, and run sh inference.sh on WAV files placed in noisy_audio/. Cleaned audio lands in enhanced_audio/. The model has only 9.6 million parameters, and a creator demonstrated it running on a CPU with the GPU idle. NVIDIA's model card, however, lists GPU hardware as the tested target, so verify CPU behavior on your own machine.

What is NVIDIA NemotronLabs RE-USE?

RE-USE is a multilingual universal speech enhancement (USE) model. Its job is to take degraded speech and restore quality while keeping the speaker's voice intact. NVIDIA lists these degradations as covered:

• additive noise
• reverberation
• clipping
• bandwidth limitation
• codec artifacts
• packet loss
• low-quality microphones

For overlapping speech, the model is trained to keep the main speaker and suppress the others. NVIDIA also points to a separate laughter variant if you need laughter preserved better. The intended users are content creators who want cleaner recordings and researchers who want to polish speech datasets before training TTS or speech LLMs.

How do you install and run it?

The model card describes the whole flow. Mamba itself can be fiddly to set up, so the card links pre-built Docker environments (from the SEMamba project) to simplify that part.

pip install resampy

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

sh inference.sh

Resampy is needed only for bandwidth extension, which you control through the BWE argument in the script. Put your noisy files in noisy_audio/ first. Input and output are both mono WAV.

What about long files?

The training clips were all under 10 seconds. For files longer than about 30 seconds, or if you hit out-of-memory errors, use the chunked script instead:

sh inference_chunk.sh

Drop files in long_noisy_audio/ and collect results from Long_enhanced_audio/. You can tune chunk_size_in_seconds and hop_length_portion inside the script.

Does it really run on a CPU?

A creator showed it running locally with the GPU idle, and a model this small makes that plausible. But the model card says NVIDIA's models are designed and optimized for GPU-accelerated systems. It lists NVIDIA Ampere as the supported microarchitecture, Linux as the supported OS, Triton as the acceleration engine, and an A100 as the hardware requirement.

Treat CPU use as something that worked in a demo, not as a supported configuration. No CPU speed or real-time-factor figures appear in the sources, so benchmark your own files before planning a pipeline around it. A quick test on a few minutes of audio will tell you more than any spec sheet.

How well does it clean speech in other languages?

The creator ran clips in several languages through the hosted demo. In a Hindi clip, traffic noise at the start and a clicking sound near the end both disappeared. Another language sample had noise at the start, and the enhanced version removed it.

That fits the card's claim of language-agnostic behavior. The training data was about 2,500 hours, weighted toward English, with German, Spanish, French and Mandarin (zh-CN) from Multilingual LibriSpeech and Common Voice. Hindi is not in the listed training sets, which makes that clip a useful sign of generalization. It is still one anecdotal sample, not a measured result.

How does it score against other models?

NVIDIA evaluated on subsets of the URGENT 2025 non-blind test set against clean reference speech. On the full set:

RE-USE leads on NISQA, PESQ, STOI, speaker similarity and ASR character accuracy. It trails Sidon on DNSMOS and UTMOS, and UniPASE edges it on SpeechBERTScore. So the claim is not that it wins everything. It is that a model roughly 25 to 50 times smaller competes closely, and it beats the others on the fidelity-style metrics. The card also shows it ahead of UniSE at 16 kHz, Resemble Enhance at 44.1 kHz and ClearVoice at 48 kHz on most columns.

These are NVIDIA's own numbers on a synthetic test set, so real recordings may behave differently.

Where does it fail?

Music is the clear weak spot. In the creator's test, an AI-generated video had noise at the start and middle and music near the end. The enhanced output had no audio at all. The model treated everything as noise, which the spectrogram confirmed.

A second AI-generated video fared better. Heavy noise was removed, though a screeching sound stayed in to match the video. The creator suspected the AI-generated audio played a part but called the judgment subjective.

The cause is plausible from the design. The model is trained for speech enhancement, with noise drawn from AudioSet and FreeSound and speech-only clean targets. Music was not framed as something to preserve. If your material mixes speech with a score or sound effects, plan to separate the voice track first or expect the music to go.

How does the architecture work?

The model uses a convolutional encoder and decoder with a bidirectional Mamba stack in the middle, 30 layers deep, for time-frequency modeling. The convolutional parts handle turning the signal into features and rebuilding audio at the end. The Mamba layers model how speech changes across time and frequency. Mamba is a state space architecture, so it scales more gently with sequence length than attention does, which helps explain how the whole thing fits in 9.6M parameters.

Can you use it commercially?

NVIDIA states the model is ready for commercial use, released under the OpenMDW License Agreement 1.1. Read the license text yourself before building a product on it. The repository is also not accepting contributions.