Audio-to-Audio Index
Speech enhancement, restoration and audio editing models on the Hugging Face Hub, grouped by what they do. Curated by Daniel Rosehill; source collections: Voice Enhancement and Audio editing utilities.
27 entries · Collections: Voice Enhancement · Audio editing utilities
Universal enhancement and restoration
One model for noise, reverb, codec damage and band-limiting. Mostly generative; best quality, not real-time.
NVIDIA's current RE-USE release: universal speech enhancement (Mamba), any sampling rate, language-agnostic.
modelRE-USEnvidiaThe original RE-USE release.
modelstupasecisco-aiStuPASE: generative enhancement aiming at studio quality; removes noise and reverb, keeps speaker identity. 16 kHz.
modelsidon-v0.1sarulab-speechSidon: speech restoration to clean studio-like audio.
modelVoiceRestorejadechoghariVoiceRestore: flow-matching transformer for degraded voice recordings.
modelresemble-enhanceResembleAIResemble Enhance: denoiser plus generative enhancer; widely used.
modeldiamond-1.0nineninesixCodec-based autoregressive speech restoration.
modelDAC-SE1disco-ethDAC-SE1: enhancement as sequence modelling over discrete audio tokens.
modelspeech-enhancement-sgmsesp-uhhSGMSE+: diffusion-based enhancement and dereverberation; research reference.
Dialogue and speaker-aware
Clean speech while separating or suppressing other speakers.
Real-time and on-device denoising
Small or streaming models for calls, live audio and phones.
Streaming RE-USE.
modelcleardesert-ant-labsLightweight denoise and dereverb; ONNX and TFLite builds.
modelLocalVQELocalAI-ioNoise suppression plus acoustic echo cancellation; GGUF.
modelDeepFilterNet3-ONNXsoniqoDeepFilterNet3 as ONNX (upstream lives on GitHub).
modelMP-SENet-DNSJacobLinCoolMP-SENet trained on DNS Challenge data.
modelMossFormer2-SEstarkdmiMossFormer2 enhancement, MLX port for Apple silicon.
modelmetricgan-plus-voicebankspeechbrainMetricGAN+ via SpeechBrain; classic baseline.
Bandwidth extension and upscaling
Recover high frequencies from low-rate or compressed audio.
Ready-to-run builds
GGUF conversions for CPU runtimes.
Instruction-guided audio editing
Edit audio from a text prompt: add, remove or replace sounds.