Studio
Browse API endpoints
post/upscale-audio

Restore low-quality audio and export a cleaner WAV

Send a compressed song, damaged recording, voice track, or stem. Choose what needs fixing. Your app or agent gets a restored WAV that is ready for editing, mastering, or download.

Get an API keyView examples

How this endpoint works

Use this API when users need to upscale MP3 to WAV, restore missing high frequencies, fix compressed AI music, clean speech, remove clipping, reduce reverb, or repair a stem. Unlike file conversion, restoration predicts missing audio content. Choose Apollo, AudioSR, UniverSR, or a problem-specific preset after you know what the source needs.

  1. 1

    Choose the audio and problem

    Send the track or stem ID and choose the restoration preset that matches the damage.

  2. 2

    Wait for restoration

    Save id and poll the restoration job until it finishes or fails.

  3. 3

    Download the restored WAV

    Download the cleaner file or send it directly to mastering or stem separation.

Common use cases

Choose this operation when it matches the source and result your pipeline needs.

Upscale MP3 to WAV

Rebuild missing high frequencies and reduce MP3 compression artifacts before editing or mastering.

Fix hiss, clipping, or reverb

Choose a problem-specific preset for a noisy AI song, clipped recording, reverberant vocal, or damaged stem.

Restore voice recordings

Improve low-bandwidth speech, vocals, podcasts, phone calls, or voice notes before publishing.

Processing used by 80,000+ music makers

Results people rely on

“I've been loving the app. The audio restoration works amazing on my suno songs”

Jordi

Audio Super Resolution

“Amazing tool for audio. Clean, simple, and effective. I would spend hours in RX to get the same results. Give it a try. I can save you hours of production time.”

Dan Campbell ~Riffster

“It can bring back clarity without making everything sound too artificial or overprocessed. Stem separation is also surprisingly clean.”
R

R.

Audio restoration and stem separation

“Does amazing job in restoring Audio from bad quality MP3s and badly extracted audio from CDs. Always my go to for audio restoration.”
N

N.

MP3 and CD audio restoration

Hear audio restoration results

These shared user results demonstrate the restoration output available through the same processing system.

More User Examples

Choose what you want to fix

MP3 Music Restoration (Apollo, 2025)

preset="universal_enhancer"

This model was trained to regenerate high-quality music from its low-quality MP3 version. It fills gaps in the spectrogram and restores missing frequencies above 16 kHz. Perfect for your old 128 kbps MP3s downloaded from the internet and lost media.

Parameters
stereo_modeStereo Modestring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.

AI Cover - Stable Audio 3 (Stable Audio 3 Medium, 2026)

preset="stable_audio_3"

This instrumental music model generates an AI cover using the input audio as a reference. The settings let you make the result more or less similar to the original. Does not work with singing! Remix your instrumentals, turn noise into musical sounds, and repair low frequencies.

Parameters
model_nameModel variantstring | null

Selects the model variant used inside the restoration preset. For denoise and decrowd, sam_audio requires an explicit prompt. Also supported by declip (simple_declip or betterdeclipper), dereverb, dialogue_isolate, universr, ap_bwe, acestep_15_xl, and stable_audio_3. UniverSR variants support audio/vocal super-resolution, AP-BWE variants output 48 kHz (ap-bwe-48k, default) or 16 kHz (ap-bwe-16k) speech, and ACE-Step/Stable Audio variants support prompt-guided remastering. Other presets choose their model from preset and ignore this field.

Allowed values

  • stable-audio-3-mediumStable Audio 3 MediumRecommended
strengthRemix Strengthnumber

High values produce results very different from the input audio. Sometimes too different.

Common values

  • 0.01Minimal edit
  • 0.1Light edit
  • 0.3Bit more edit
  • 0.5Balanced editRecommended
  • 0.6Moderate remix
  • 0.7Strong remix
  • 0.95Totally different songDefault
promptPromptstring

Use keywords to describe the style you want. For example: *italian western, cinematic score, mellow strings and choir*

Default"high quality studio recording CD quality"

Example

"clean studio master, full bandwidth, natural transients"
prompt_strengthPrompt Strengthnumber

Increases the prompt's influence relative to the original music. High values produce the most creative and different results.

Range: 0 to 10

Common values

  • 0.5Less prompt strengthThe prompt is mostly ignored.
  • 1Base prompt strengthDefaultRecommendedBalanced.
  • 2More promptA little more prompt.
  • 3Stronger promptEven more prompt.
  • 4Even more promptOK, this is where it gets really fun. The model responds much better to the prompt.
  • 5High prompt strengthCareful: every comma in the prompt counts!
  • 6Very high prompt strengthWow! You might hear some entirely new sounds.
  • 7Maximum promptAt this level, you can use any input audio: the model follows only the prompt.
inpaint_regionsInpaint regionsArray<{ start: number; end: number }> | null

Lists source time ranges to regenerate while preserving the rest of the input audio. Each range requires start and end times in seconds. Supported by stable_audio_3 only; other presets ignore this field. Omit it to run an ordinary audio-to-audio remix.

Example

[{"end":8,"start":4}]

AI Cover - YuE 2 (YuE2 3B, 2026)

preset="yue_2"

This music model detects melodies, harmonies, and sometimes even lyrics in the reference audio. It uses them to generate a high-quality cover. The result usually has a very different rhythm. The model loves to sing gibberish. Treat it as creative raw material.

Parameters
promptPromptstring

Use keywords to describe the genre, instruments, language, voice...

Default"high quality studio recording CD quality"

Example

"clean studio master, full bandwidth, natural transients"
instrumental_onlyInstrumentalboolean

Combines the detected vocal and instrumental melodies and gives the model no input for the vocal track. Unfortunately, the model usually adds gibberish vocals anyway.

Defaultfalse

prompt_strengthPrompt Strengthnumber

How strongly the cover follows your style prompt.

Range: 0 to 10

Common values

  • 0.1Minimal
  • 0.5Light
  • 11.0 - StandardDefaultRecommended
  • 1.2Strong

AI Cover - ACEStep 1.5 XL (ACE-Step 1.5 XL, 2026)

preset="acestep_15_xl"

This instrumental music model generates an AI cover using the input audio as a reference. Generations often sound very close to the reference unless you greatly reduce the original's influence. Recommended for improving bass tracks, reconstructing missing notes, and correcting specific artifacts.

Parameters
model_nameModelstring | null

Selects the model variant used inside the restoration preset. For denoise and decrowd, sam_audio requires an explicit prompt. Also supported by declip (simple_declip or betterdeclipper), dereverb, dialogue_isolate, universr, ap_bwe, acestep_15_xl, and stable_audio_3. UniverSR variants support audio/vocal super-resolution, AP-BWE variants output 48 kHz (ap-bwe-48k, default) or 16 kHz (ap-bwe-16k) speech, and ACE-Step/Stable Audio variants support prompt-guided remastering. Other presets choose their model from preset and ignore this field.

Allowed values

  • acestep-v15-xl-turboAceSTEP turbo
  • acestep-v15-xl-sftAceSTEP SFT
strengthOriginal Influencenumber

How much the original audio influences the result. 1.0 means the model tries to match the original as close as possible. 0.1 means it's a loose reference.

Common values

  • 1.0As close as possible
  • 0.99Almost original
  • 0.5Inspiration
  • 0.1Creative interpretation
promptPromptstring

Use keywords to describe the style, instruments, tempo, and key.

Default"high quality studio recording CD quality"

Example

"clean studio master, full bandwidth, natural transients"

High-Frequency Restoration - AudioSR (Audio SR, 2022)

preset="audiosr"

This model regenerates high frequencies above the selected cutoff and keeps the original audio below it. The lower the cutoff, the more of the sound it reconstructs. Great for historical recordings, poor-quality recordings, or recordings with unwanted high-frequency noise. It does not change the bass (use an AI cover model instead).

Parameters
frequency_cutoffStrengthinteger

Sets the frequency above which AudioSR regenerates the high end. A low cutoff reconstructs more of the frequency spectrum. A high cutoff focuses on the highest frequencies (which are less audible).

Allowed values

  • 16000Very Subtle (16 kHz cutoff)
  • 13000Subtle (13 kHz cutoff)Default
  • 10000Mild (10 kHz cutoff)
  • 8000Medium (8 kHz cutoff)Recommended
  • 5000Pronounced (5 kHz cutoff)
  • 4000Strong (4 kHz cutoff)
  • 3000Very Strong (3 kHz cutoff)
reconstruction_methodReconstructionstring

Controls how generated high frequencies are combined with the source audio. Supported by audiosr and universr. For AudioSR, multiband_ensemble low-passes the original audio at frequency_cutoff minus 1000 Hz, high-passes the AudioSR output at the same crossover, then sums both bands. original_signal uses frequency_cutoff as a hard final spectrum boundary: original source bins below the cutoff and generated bins at or above it. For UniverSR, original preserves the legacy reconstruction path, while original_signal keeps the bandwidth-limited input for model conditioning but takes the final low-frequency bins from the original 48 kHz source signal. seamless_reconstruction_soft uses the same source low band and caps each generated frequency bin's energy at the maximum source energy in the 500 Hz below frequency_cutoff, per time frame and processed channel. Other presets ignore this field.

Allowed values

  • multiband_ensembleMultiband EnsembleRecommendedA gradual transition between the original low frequencies and the restored high frequencies.
  • original_signalContinuous ReconstructionA sharp cutoff between the original audio below the threshold and the restored high frequencies.
stereo_modeChannelsstring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.

High-Frequency Restoration - UniverSR (UniverSR, 2026)

preset="universr"

An upgraded version of AudioSR for music, voice, and sound effects. UniverSR reconstructs missing or poor-sounding high frequencies to add brightness. Try it on dull voices, old recordings (such as cassettes), or AI music from SUNO. It does not change the low frequencies (use an AI cover model instead).

Parameters
frequency_cutoffStrengthinteger

Sets the upper frequency boundary in hertz where a bandwidth-extension model starts rebuilding audio. Supported by audiosr and universr, in 100 Hz steps up to 22000. AudioSR accepts 3000 or higher. UniverSR accepts 4000 or higher; it runs its model at the highest of 4000, 6000, 8000, or 12000 that does not exceed the cutoff and keeps generated audio only above the cutoff. AP-BWE accepts 4000, 6000, 8000, or 12000 with ap-bwe-48k (default 8000), and 1000, 2000, or 4000 with ap-bwe-16k (default 4000); each cutoff selects the model trained on audio band-limited to it. Other presets ignore this field.

Allowed values

  • 12000Subtle (12 kHz cutoff)
  • 16000Very Subtle (16 kHz cutoff)
  • 8000Medium (8 kHz cutoff)Recommended
  • 6000Pronounced (6 kHz cutoff)
  • 4000Strong (4 kHz cutoff)
reconstruction_methodReconstructionstring

Controls how generated high frequencies are combined with the source audio. Supported by audiosr and universr. For AudioSR, multiband_ensemble low-passes the original audio at frequency_cutoff minus 1000 Hz, high-passes the AudioSR output at the same crossover, then sums both bands. original_signal uses frequency_cutoff as a hard final spectrum boundary: original source bins below the cutoff and generated bins at or above it. For UniverSR, original preserves the legacy reconstruction path, while original_signal keeps the bandwidth-limited input for model conditioning but takes the final low-frequency bins from the original 48 kHz source signal. seamless_reconstruction_soft uses the same source low band and caps each generated frequency bin's energy at the maximum source energy in the 500 Hz below frequency_cutoff, per time frame and processed channel. Other presets ignore this field.

Allowed values

  • seamless_reconstruction_softSoft ReconstructionRecommendedRecommended. Attenuates the restored high frequencies for a softer sound. Fewer audible artifacts, but less brightness. Boost it with an EQ.
  • original_signalContinuous ReconstructionBlends restored high frequencies smoothly with the original audio.
  • originalAuthors' TechniqueDefaultThe original method from the research paper. This technique often leaves a visible transition in the spectrogram.
model_nameModelstring | null

Selects the model variant used inside the restoration preset. For denoise and decrowd, sam_audio requires an explicit prompt. Also supported by declip (simple_declip or betterdeclipper), dereverb, dialogue_isolate, universr, ap_bwe, acestep_15_xl, and stable_audio_3. UniverSR variants support audio/vocal super-resolution, AP-BWE variants output 48 kHz (ap-bwe-48k, default) or 16 kHz (ap-bwe-16k) speech, and ACE-Step/Stable Audio variants support prompt-guided remastering. Other presets choose their model from preset and ignore this field.

Allowed values

  • universr-audioOriginalRecommended
  • universr-audio-finetune-v1Fine-tune v1
  • pitbull_edm_206Pitbull EDM
stereo_modeStereostring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)RecommendedProcess the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.

Speech Enhancement - NovaSR Upscaler (NovaSR, 2026)

preset="novasr"

This upscaling model was trained on English speech. Use it to remove the walkie-talkie sound from your podcasts, voice-overs, or isolated vocal tracks.

Parameters
stereo_modeChannelsstring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.

High-Frequency Restoration - FlashSR (FlashSR, Jan 2025)

preset="flashsr"

Similar to AudioSR, this model reconstructs high frequencies more quickly. Handy for quickly upscaling audio from 44.1 kHz to 48 kHz. It does not change the low frequencies (use an AI cover model instead).

Parameters
stereo_modeStereo Modestring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.

Correct Tempo Variations (Constant BPM) (Beat This!, 2024)

preset="constant_bpm"

SUNO tracks often have an unstable tempo, subtly speeding up or slowing down throughout the music. This model detects the tempo and corrects these variations. Ideal before importing a track into a DAW or playing it in a DJ set. Also useful for live recordings. Works best when there are drums: EDM, rock, pop, techno, house, etc.

Parameters
quantize_precisionBeat quantizationstring

Snap detected beats to the nearest musical subdivision.

Allowed values

  • 1/4
  • 1/8
  • 1/8T
  • 1/16DefaultRecommended
  • 1/16T
  • 1/32
  • 1/32T

Neural Reconstruction (DACVAE, 2024)

preset="dacvae"

This model passes the sound through a high-quality encoder-decoder model. The effect is very subtle: it can help correct certain artifacts (especially hi-hats in your drum'n'bass tracks).

Parameters
stereo_modeChannelsstring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.

Correct Clipping (Declip)

preset="declip"

Heavily compressed or limited audio sometimes has a waveform that is flat in places (clipping). This digital artifact, caused during recording or mixing, creates unpleasant noise. Use this model to correct it.

Parameters
model_nameModelstring | null

Selects the model variant used inside the restoration preset. For denoise and decrowd, sam_audio requires an explicit prompt. Also supported by declip (simple_declip or betterdeclipper), dereverb, dialogue_isolate, universr, ap_bwe, acestep_15_xl, and stable_audio_3. UniverSR variants support audio/vocal super-resolution, AP-BWE variants output 48 kHz (ap-bwe-48k, default) or 16 kHz (ap-bwe-16k) speech, and ACE-Step/Stable Audio variants support prompt-guided remastering. Other presets choose their model from preset and ignore this field.

Allowed values

  • simple_declipSimple CorrectionRecommendedAutomatically corrects clipping while maintaining high perceived loudness.
  • betterdeclipperAdvanced CorrectionSeveral experimental mathematical techniques to remove clipping from very, very damaged audio.
modeModestring

Auto mode is recommended.

Allowed values

  • autoAutoDefaultRecommendedAnalyzes flat clipping, smeared peaks from lossy encoding or resampling, limiter ceilings, and soft shoulders. Restores the detected damage and leaves unaffected regions intact.
  • hardHard clippingRestores detected clipping ceilings without a soft shoulder. At limiter ceilings, repairs only flat runs of two or more samples.
  • softSoft shoulderUses the same analysis as Auto but always restores a soft region below each ceiling, using the detected knee or 80% of the ceiling.
  • limiterLimiter (experimental)Restores a deeper soft region from 50% of limiter ceilings. Can reshape clean peaks and add audible distortion; compare carefully with Auto.
  • legacyLegacyUses the previous amplitude-histogram detection: repairs flat clipping plateaus, or a soft knee when no plateau is found.

Remove Limiting (Delimiter) (BS-RoFormer De-limit, becruily)

preset="delimiter"

Undo the brickwall limiting of an over-compressed master and bring back its punch and dynamics. In 16 or 24-bit, the result is 12 dB quieter to leave room for the restored peaks. Choose 32-bit float in the settings to keep the original level.

Noise Reduction

preset="denoise"

Remove hiss, hum, background noise, rumble, or animal and crowd sounds. Choose the variant based on the noise present and the sound you want to keep.

Parameters
model_nameModelstring | null

Selects the model variant used inside the restoration preset. For denoise and decrowd, sam_audio requires an explicit prompt. Also supported by declip (simple_declip or betterdeclipper), dereverb, dialogue_isolate, universr, ap_bwe, acestep_15_xl, and stable_audio_3. UniverSR variants support audio/vocal super-resolution, AP-BWE variants output 48 kHz (ap-bwe-48k, default) or 16 kHz (ap-bwe-16k) speech, and ACE-Step/Stable Audio variants support prompt-guided remastering. Other presets choose their model from preset and ignore this field.

Allowed values

  • denoiseDenoiseRecommended
  • denoise_debleedDenoise and Debleed
  • custom_speechSpeech only
  • custom_music_singingMusic with singing
  • custom_musicMusic Only (no vocals)
  • custom_singingSinging only (no instruments)
promptSound to keepstring

Describe the sound to keep in the cleaned track.

Default"high quality studio recording CD quality"

Example

"clean studio master, full bandwidth, natural transients"
stereo_modeStereo modestring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.
spansSound hintsArray<{ start: number; end: number }> | null

Optional time hints for SAM Audio denoise and decrowd variants.

Remove Reverb

preset="dereverb"

Get a dry sound from a recording drenched in reverb or echo, such as one recorded in a cellar or a church. Also useful for removing the slight echo of a poorly soundproofed room.

Parameters
model_nameModelstring | null

Selects the model variant used inside the restoration preset. For denoise and decrowd, sam_audio requires an explicit prompt. Also supported by declip (simple_declip or betterdeclipper), dereverb, dialogue_isolate, universr, ap_bwe, acestep_15_xl, and stable_audio_3. UniverSR variants support audio/vocal super-resolution, AP-BWE variants output 48 kHz (ap-bwe-48k, default) or 16 kHz (ap-bwe-16k) speech, and ACE-Step/Stable Audio variants support prompt-guided remastering. Other presets choose their model from preset and ignore this field.

Allowed values

  • dereverbLong Reverb (strong)Recommended
  • dialogue_isolateRoom echo (light)

Crowd Noise Reduction

preset="decrowd"

Improve your live recordings! Keep the performance and remove audience noise: coughing, whistling, shouting, applause, etc. Use the Before/After slider to control the ambience level.

Parameters
model_nameModelstring | null

Selects the model variant used inside the restoration preset. For denoise and decrowd, sam_audio requires an explicit prompt. Also supported by declip (simple_declip or betterdeclipper), dereverb, dialogue_isolate, universr, ap_bwe, acestep_15_xl, and stable_audio_3. UniverSR variants support audio/vocal super-resolution, AP-BWE variants output 48 kHz (ap-bwe-48k, default) or 16 kHz (ap-bwe-16k) speech, and ACE-Step/Stable Audio variants support prompt-guided remastering. Other presets choose their model from preset and ignore this field.

Allowed values

  • decrowdRoformerRecommended
  • decrowd_mdxMDX Net
  • custom_music_singingSAM Audio
promptSound to keepstring

Describe the sound to keep in the cleaned track.

Default"high quality studio recording CD quality"

Example

"clean studio master, full bandwidth, natural transients"
stereo_modeStereo modestring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.
spansSound hintsArray<{ start: number; end: number }> | null

Optional time hints for SAM Audio denoise and decrowd variants.

Keep Only Center Mono (Gillian CenterWide v3, 2026)

preset="phantom_center"

This model keeps the mono center of a track and removes the content spread across the sides. Handy for isolating a bass, kick drum, or podcast voice. You can also use it to remove phaser, chorus, or flanger from an instrument.

Stereoize (Gillian Stereoizer Alpha 1)

preset="stereoize"

Convert mono audio into stereo by generating separate left and right channels. Stereo input is mixed to mono before processing.

Speech Enhancement - RE-USE Upscaler (NVIDIA RE-USE, 2026)

preset="reuse"

This model dramatically improves speech quality in multiple languages. It recovers high-frequency detail and reduces noise, reverb, and audio defects. For the most difficult restorations of podcasts, voice messages, or singing. Hallucinations are common: use UniverSR or blend with the original audio to improve the result.

Parameters
stereo_modeChannelsstring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.

Speech Bandwidth Extension - AP-BWE (AP-BWE, 2024)

preset="ap_bwe"

Makes muffled phone calls, voice messages, and low-quality voice recordings sound clear and full again.

Parameters
model_nameModelstring | null

Selects the model variant used inside the restoration preset. For denoise and decrowd, sam_audio requires an explicit prompt. Also supported by declip (simple_declip or betterdeclipper), dereverb, dialogue_isolate, universr, ap_bwe, acestep_15_xl, and stable_audio_3. UniverSR variants support audio/vocal super-resolution, AP-BWE variants output 48 kHz (ap-bwe-48k, default) or 16 kHz (ap-bwe-16k) speech, and ACE-Step/Stable Audio variants support prompt-guided remastering. Other presets choose their model from preset and ignore this field.

Allowed values

  • ap-bwe-48k48 kHz upsamplingRecommended
  • ap-bwe-16k16 kHz upsampling
frequency_cutoffStrengthinteger

Sets the frequency above which AP-BWE rebuilds the voice. Pick the frequency where your recording's high end stops: a lower cutoff rebuilds more of the spectrum.

Allowed values

  • 12000Subtle (12 kHz cutoff)
  • 8000Medium (8 kHz cutoff)Recommended
  • 6000Pronounced (6 kHz cutoff)
  • 4000Strong (4 kHz cutoff)
frequency_cutoffStrengthinteger

Sets the frequency above which AP-BWE rebuilds the voice. Pick the frequency where your recording's high end stops: a lower cutoff rebuilds more of the spectrum.

Allowed values

  • 4000Strong (4 kHz cutoff)Recommended
  • 2000Very Strong (2 kHz cutoff)
  • 1000Extreme (1 kHz cutoff)
stereo_modeChannelsstring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.

MP3 Singing Restoration (Apollo Voice, 2025)

preset="apollo_voice"

An MP3 restoration model specialized in singing. Handy for improving a cappella vocals downloaded as MP3s from the internet or correcting certain compression artifacts.

Parameters
stereo_modeStereo Modestring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

Allowed values

  • mid_sidesStereo (Mid/Sides)DefaultRecommendedProcess Mid ((L + R) / 2) and Sides ((L - R) / 2) independently, then recombine them as stereo.
  • sides_onlySides onlyProcess the sides and keep the original center.
  • mid_onlyMid onlyProcess the center and keep the original sides.
  • left_rightStereo (L/R)Process the Left and Right channels independently.
  • monoMonoAverage Left and Right into one mono channel, then process it once.

Deeffect Vocals (Gillian DeEffect v1 beta, 2026)

preset="deeffect_vocals"

Get a clean, dry lead vocal from a processed recording. Removes reverb, delay, vocal doubles, and backing vocals, and suppresses the instruments.

Code examples

Server-side example

JavaScript (Node.js)
import { writeFile } from "node:fs/promises";

const API_URL = "https://api.neuralanalog.com";
const API_KEY = process.env.NEURALANALOG_API_KEY;

async function waitForCompletion(objectType, objectId) {
  while (true) {
    const response = await fetch(`${API_URL}/status/${objectType}/${objectId}`, {
      headers: { "X-API-Key": API_KEY },
    });
    if (!response.ok) {
      throw new Error(`Status check failed with ${response.status}`);
    }

    const status = await response.json();
    if (status.is_failed) {
      throw new Error(status.error_message || `${objectType} processing failed`);
    }
    if (status.is_complete) {
      return status;
    }
    await new Promise((resolve) => setTimeout(resolve, 5000));
  }
}

const requestResponse = await fetch(`${API_URL}/upscale-audio`, {
  method: "POST",
  headers: {
    "X-API-Key": API_KEY,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
  "audio_id": "6c62f8e7-02a3-48c0-a5b5-5de87ed9c31a",
  "preset": "universal_enhancer",
  "bit_depth": 24
}),
});
if (!requestResponse.ok) {
  throw new Error(`Restoration request failed with ${requestResponse.status}`);
}

const result = await requestResponse.json();
const jobId = result["id"];
const status = await waitForCompletion("upscaled", jobId);

const downloadResponse = await fetch(
  `${API_URL}/download/upscaled/${jobId}`,
  { headers: { "X-API-Key": API_KEY }, redirect: "follow" },
);
if (!downloadResponse.ok) {
  throw new Error(`Download failed with ${downloadResponse.status}`);
}
await writeFile("restored.wav", Buffer.from(await downloadResponse.arrayBuffer()));

Parameters

Send the API key from a trusted server. Never expose it in client-side JavaScript.

JSON request body

audio_id
requiredstring

Identifies the full audio file to restore.

Example: "6c62f8e7-02a3-48c0-a5b5-5de87ed9c31a"

preset
optionalstring

Selects the restoration model or repair process to run. Each preset supports its own parameters; fields that do not apply to the selected preset are ignored.

"universal_enhancer""constant_bpm""novasr""flashsr""dacvae""declip""delimiter""dialogue_isolate""denoise""dereverb""decrowd""phantom_center""stereoize""audiosr""apollo_voice""deeffect_vocals""lavasr""reuse""ap_bwe""acestep_15_xl""stable_audio_3""yue_2""universr"

Default: "universal_enhancer"

stereo_mode
optionalstring

Controls how stereo material is processed. Defaults to mid_sides, which processes center and side content separately, except for universr, which defaults to left_right; left_right processes channels independently, mono folds to mono, and sides_only/mid_only process one component while preserving the other. Supported by universal_enhancer, apollo_voice, universr, reuse, ap_bwe, flashsr, audiosr, novasr, dacvae, and lavasr. Other presets ignore this field.

"mid_sides""left_right""mono""sides_only""mid_only"

Default: "mid_sides"

quantize_precision
optionalstring

Snaps detected beats to this musical subdivision when using constant_bpm. Other restoration presets ignore this field.

"1/4""1/8""1/8T""1/16""1/16T""1/32""1/32T"

Default: "1/16"

frequency_cutoff
optionalinteger

Sets the upper frequency boundary in hertz where a bandwidth-extension model starts rebuilding audio. Supported by audiosr and universr, in 100 Hz steps up to 22000. AudioSR accepts 3000 or higher. UniverSR accepts 4000 or higher; it runs its model at the highest of 4000, 6000, 8000, or 12000 that does not exceed the cutoff and keeps generated audio only above the cutoff. AP-BWE accepts 4000, 6000, 8000, or 12000 with ap-bwe-48k (default 8000), and 1000, 2000, or 4000 with ap-bwe-16k (default 4000); each cutoff selects the model trained on audio band-limited to it. Other presets ignore this field.

Default: 13000

model_name
optionalstring | null

Selects the model variant used inside the restoration preset. For denoise and decrowd, sam_audio requires an explicit prompt. Also supported by declip (simple_declip or betterdeclipper), dereverb, dialogue_isolate, universr, ap_bwe, acestep_15_xl, and stable_audio_3. UniverSR variants support audio/vocal super-resolution, AP-BWE variants output 48 kHz (ap-bwe-48k, default) or 16 kHz (ap-bwe-16k) speech, and ACE-Step/Stable Audio variants support prompt-guided remastering. Other presets choose their model from preset and ignore this field.

"simple_declip""betterdeclipper""dereverb""dialogue_isolate""denoise""denoise_debleed""decrowd""decrowd_mdx""sam_audio""universr-audio""universr-audio-finetune-v1""universr-vocal""pitbull_edm_206""acestep-v15-xl-turbo""acestep-v15-xl-sft""stable-audio-3-medium""ap-bwe-48k""ap-bwe-16k"
mode
optionalstring

Better Declipper mode: auto analyzes flat or smeared clipping, limiter ceilings, and soft shoulders; hard restores ceilings without a soft region; soft always restores a soft shoulder below detected ceilings; limiter restores a deeper soft region and can add distortion (experimental); legacy uses the previous plateau-or-knee detection. Other restoration models ignore this field.

"auto""hard""soft""limiter""legacy"

Default: "auto"

reconstruction_method
optionalstring

Controls how generated high frequencies are combined with the source audio. Supported by audiosr and universr. For AudioSR, multiband_ensemble low-passes the original audio at frequency_cutoff minus 1000 Hz, high-passes the AudioSR output at the same crossover, then sums both bands. original_signal uses frequency_cutoff as a hard final spectrum boundary: original source bins below the cutoff and generated bins at or above it. For UniverSR, original preserves the legacy reconstruction path, while original_signal keeps the bandwidth-limited input for model conditioning but takes the final low-frequency bins from the original 48 kHz source signal. seamless_reconstruction_soft uses the same source low band and caps each generated frequency bin's energy at the maximum source energy in the 500 Hz below frequency_cutoff, per time frame and processed channel. Other presets ignore this field.

"multiband_ensemble""original""original_signal""seamless_reconstruction_soft"

Default: "original"

strength
optionalnumber

Controls processing intensity from subtle cleanup to aggressive restoration. Higher values preserve less of the degraded source. Supported by acestep_15_xl and stable_audio_3. Other presets ignore this field.

Default: 0.95

prompt
optionalstring

Describes the sound that a prompt-guided model should create from the source audio. Supported by acestep_15_xl, stable_audio_3, and yue_2. Other presets ignore this field. Keep the prompt under 256 characters.

Default: "high quality studio recording CD quality"

Example: "clean studio master, full bandwidth, natural transients"

prompt_strength
optionalnumber

Controls how strongly Stable Audio 3 or YuE 2 follows the text prompt. Higher values give the prompt more influence. YuE 2 maps this to cfg_scale while retaining its melody score conditioning. Supported by stable_audio_3 and yue_2; other presets ignore this field. Accepts values from 0 to 10.

Default: 1

instrumental_only
optionalboolean

For YuE 2, merge the vocal and instrumental melodies into one instrumental part and request generation without vocals. Vocal notes take priority where melodies overlap. Other presets ignore this field.

Default: false

inpaint_regions
optionalarray<object> | null

Lists source time ranges to regenerate while preserving the rest of the input audio. Each range requires start and end times in seconds. Supported by stable_audio_3 only; other presets ignore this field. Omit it to run an ordinary audio-to-audio remix.

Example: [{"end":8,"start":4}]

spans
optionalarray<object> | null

Optional time hints for SAM Audio denoise and decrowd variants.

source_effects_applied
optionalboolean | null

Whether the rendered source includes FX; false preserves live FX when selecting the output.

stem_id
optionalstring | null

Selects one stem from the parent audio as the restoration source. Omit it to restore the full audio file.

Example: "abf8a992-1c4e-4935-93f0-197116e77e49"

source_upscaled_id
optionalstring | null

Selects an existing restored version as the source for another pass.

Example: "d66cf940-bf26-45bb-80f7-332f26b6859a"

source_mastered_id
optionalstring | null

Selects an existing mastered artifact as the restoration source.

Example: "f5db8e4b-2e74-4198-a8de-0c3a398620e9"

source_transcription_id
optionalstring | null

Source transcription: a YuE 2 melody, or lineage for a rendered temporary mix.

source_temporary_mix_key
optionalstring | null

Short-lived Current Main Mix or Current All Stems Mix R2 source key.

selection
optionalobject | null

Limits restoration to a start and end time in seconds. Omit it to restore the full selected source.

Example: {"end":42,"start":12.5}

bit_depth
optionalinteger

Output WAV bit depth for the restored audio; 32 is float.

162432

Default: 24

hq_streaming_format
optionalstring

Selects the compressed format generated for browser playback and streaming alongside the restored WAV output.

"aac""mp3""flac"

Default: "aac"

Successful response

200Successful Response
id
requiredstring

ID of the queued restored audio version.

Example: "d66cf940-bf26-45bb-80f7-332f26b6859a"

status
requiredstring

Queueing status for the restoration job.

Example: "processing"

message
requiredstring

Human-readable queueing result.

Example: "Audio restoration queued"

Errors

A missing or invalid X-API-Key returns an authentication error. Validation errors use the declared 422 response below.
422Validation Error
detail
optionalarray<object>

No description provided.

Next steps

pollDownloadMaster the restored audioTry audio restorationRead the restoration guide