Neural Analog Restoration Models

Upscale low quality MP3 back to high quality. Trained on pairs of high quality music, and their degraded mp3 versions. Restores missing high frequencies and removes compression artifacts. Output sample rate matches the prepared input rate (44.1kHz or 48kHz). Recommended for: online source imports, songs downloaded from the internet, compressed sound, tracks missing >16Khz. Not effective to regenerate 16Khz+ audio? Try AudioSR instead.

Recreates the track with Stability AI's Stable Audio 3 Medium audio-to-audio editing model. This prompt-guided remix mode uses the source audio as a reference while the text prompt steers style, tone, and instrumentation. Recommended for: creative remixes, inpainting missing musical ideas, alternate takes, and low-end or texture repair that benefits from generative reconstruction. Try a short section first for predictable results.

Recreates the track with the ACE-Step XL 2026 music generation model in remix mode. This is similar to SUNO 'cover' mode. It will use similar sounds as the reference, but use new notes if you lower the reference strength. Recommended for: Bass tracks, inpainting missing notes in stems, low end mudiness. Interesting results when blended with the original audio.

Regenerates high frequencies above the selected cutoff while keeping lower frequencies from the original audio. Output sample rate: 48kHz. Recommended for: Low quality recordings, missing high frequencies, hissing noise around 12-13Khz (hihats, sibilance, brilliance). Low thresholds (eg: 3Khz) will change audio the most, while high threshold will mostly preserve what's already here. Not recommended for: Bass tracks, muddy low end. Use Neural Remix instead.

Upscale music, voice, and sound effects to 48 kHz. UniverSR is a 2026 model developed by the University of Seoul which performs audio super-resolution directly in the complex STFT domain using flow matching. Very similar to AudioSR, but with more coherent high ends. Output sample rate: 48kHz. Recommended for: Low quality recordings, missing high frequencies, hissing noise around 12-13Khz (hihats, sibilance, brilliance). Low thresholds (eg: 4Khz) will change audio the most, while high threshold will preserve what's already here. Not recommended for: muddy bass or muddy low end. Use Neural Remix instead.

Specialized voice upscaling. Restores missing high frequencies and removes compression artifacts in lower frequencies. Output sample rate: 48kHz.

Upscaling model trained on English speech. Best for restoring podcasts, voice-overs, or isolated vocal tracks. Output sample rate: 48kHz.

Fast single-step super-resolution model for bandwidth extension and detail recovery. Output sample rate: 48kHz. Recommended for: Western music, single instrument upscaling, upscaling 44.1Khz stems to 48 kHz (before Atmos Export). Not recommended for: Bass.

Leverage the DACVAE neural codec to regenerate all frequencies. Helps with removing out of distribution frequencies. Output sample rate: 48kHz. Recommended for: weird hihats sounds in electronic music.

Fixes harsh crackling when volume is too high. Algorithms find optimal settings to remove all crackling, while still maintaining loudness.

Remove short reverb from voice recordings recorded a room with a laptop microphone. Good for podcasts, voice-overs, and dialogue that have a subtle room echo. Less effective on music or singing.

Reduces hiss, hum, and background noise while keeping the main vocals/instruments. Uses the same model family as stem splitting, but returns a single cleaned track only.

Reduces background noise and source bleed while keeping the main instrumental content. Uses the same model family as stem splitting, but returns a single cleaned track only.

Removes long reverb tails, delay, and echo to make the sound drier. Great for singing, music, and live recording with hall echo. Not too good with subtle echo in clean recording. Uses the same model family as stem splitting, but returns a single dry track only.

Removes audience noise from live recordings while preserving the performance. Uses the same model family as stem splitting, but returns a single cleaned track only.

Fast and high-quality speech upscaling. Evolution of NovaSR that uses Vocos architecture for efficiency. Output sample rate: 48kHz.

Spectral super-resolution model with selectable voice and music variants. Output sample rate depends on selected variant: Music (MUSDB)=44.1kHz, Voice 4-16=16kHz, Voice 8-16=16kHz, Voice 8-24=24kHz, Voice 12-48=48kHz. Not recommended for: Bass.