sakasegawa/NVIDIA-NemotronLabs-RE-USE-GGUF

NVIDIA NemotronLabs RE-USE GGUF for speech.cpp

This file runs only in speech.cpp 0.10.1 or later. Its layout is speech.cpp's own, so llama.cpp, LM Studio, Ollama and other GGUF readers cannot run it.

It holds NVIDIA NemotronLabs RE-USE, a SEMamba network of 9.6M parameters in F32, its STFT's settings and the chunks of the official inference_chunk.py. It takes recorded speech in any language at any rate from 8 to 48 kHz and gives back the same recording, at the same rate, for people to listen to. It removes noise, reverberation, clipping, band limits and codec damage, and with speech that overlaps it keeps the main speaker.

Use

Install speech.cpp, then:

speech enhance re-use noisy.wav -o clean.wav
speech enhance re-use --bwe 48000 phone-call.wav -o wide.wav      # resampled to 48 kHz first, to extend its band
speech asr qwen3-asr-1.7b --enhance re-use crowd.wav              # each file enhanced whole before it is recognized
speech serve re-use                                               # POST /v1/audio/enhancements

The name re-use fetches this file the first time it is used; a path to the downloaded file works as well. A program calls the C API's speech_enhance(), or its enhancements of audio given a piece at a time. speech.cpp's documentation has the model's page, the command line and the server.

Files

File	What it holds	Size	SHA-256
NVIDIA-NemotronLabs-RE-USE-10M-F32.gguf	the model in F32, the checkpoint's weights exactly	38.5 MB (38,519,136 bytes)	827faf209c6988b9cfd07b08ee7f212a7b2eeb7870a25380df671d55338b3879
NVIDIA-NemotronLabs-RE-USE-10M-F32.gguf.json	what speech info --json prints of the file
LICENSE.OpenMDW-1.1	the OpenMDW License Agreement, version 1.1

Accuracy

Each stage is checked against the official code at commit 31f82f2 on nine inputs of FLEURS speech under babble or white noise at 8 to 48 kHz. On an Apple M5 each TFMamba block lies 124.7 to 125.8 dB from the official's on the CPU and on Metal, the network 64.8 to 113.7 dB on each chunk and the enhanced audio 30.1 to 67.7 dB; on an NVIDIA GeForce RTX 2080 (Vulkan) the blocks lie 109.3 dB from it and the network 78.5 dB or more.

Enhancing each utterance before recognition lowers parakeet-tdt-0.6b-v3's WER on LibriSpeech test-clean under pub noise at 0 dB from 9.13% to 8.18%, and reazonspeech-v2's CER on FLEURS ja under crowd noise at 0 dB from 9.99% to 9.06%, but raises reazonspeech-v2's on ReazonSpeech's Japanese television from 4.69% to 8.36%: the model keeps the main speaker of speech that overlaps and removes the rest. Qwen3-ASR 1.7B, which makes fewer errors in noise to begin with, gains less: its CER on the same FLEURS ja falls from 9.33% to 9.06%, and its WER under pub noise rises a little. The model's page has every set.

Speed

The RTFx (the seconds of audio over the seconds they took) of speech enhance on 60 s of noisy speech, the median of three runs and one on the CPU, 2026-11-01:

Device	16 kHz	48 kHz
Apple M5, Metal	2.86	0.96
Apple M5, CPU	0.17	not run
NVIDIA GeForce RTX 2080, Vulkan	3.29	1.15

Each part of the audio is heard twice, as a chunk of 5 s starts every half chunk; an utterance enhanced whole, as speech asr --enhance takes it, is heard once.

Source

Converted by speech.cpp's reference/semamba/convert.py from model.safetensors (38,592,940 bytes, SHA-256 06badaf4029b297f97f7fe48d62990b48d5482608210df577354e0d4e0ed70e1) of nvidia/NVIDIA-NemotronLabs-RE-USE at revision 31f82f2efad1095034492f236e8dac9239ba713b, with the settings of its recipe and its inference_chunk.py, from the code of nvidia/NVIDIA-NemotronLabs-RE-USE at revision 31f82f2efad1095034492f236e8dac9239ba713b. Its sibling, re-use-laughter, keeps laughter, which this model removes with the noise.

License

NVIDIA NemotronLabs RE-USE is NVIDIA's, released under the OpenMDW License Agreement, version 1.1, which LICENSE.OpenMDW-1.1 holds. This file is distributed under that agreement. NVIDIA's model card has the model's ethical considerations.