ACE-Step Audio Quality: Clean Diffusion Hiss, Rumble, and Edit Clicks

An ACE-Step render can be musically convincing and still carry a thin sandy layer around cymbals, a low vibration that steals headroom, or a click where you repainted a section. None of those symptoms means the whole song is broken. It means you need to identify the exact moment, decide whether the problem is local or spread through the mix, and make the smallest useful correction before mastering.

Do not diagnose a file from the model name alone. ACE-Step can run with different model variants, inference settings, prompts, source audio, and editing workflows. A rough edge may come from generation, a later conversion, a hard edit, or an effect you added. Keep the original render, compare the same timestamp at matched loudness, and stop when the distraction is reduced without draining character from the song.

Checked September 20, 2026. The current official project is ACE-Step 1.5. Its paper describes a hybrid system in which a language model plans the song and a Diffusion Transformer, or DiT, produces the audio. The official tutorial normalizes audio-control inputs to 48 kHz stereo. The musician’s guide documents finished audio as WAV or MP3, while the official agent API tooling also exposes FLAC as an output option. Availability depends on the interface you use, so preserve the original file instead of converting a lossy download and calling it lossless.

The short answer: inspect the render before mastering

Start with three listening points: the quietest exposed passage, the busiest chorus, and every edit or repaint boundary. If the hiss moves with vocals, cymbals, or reverb, steady broadband denoise may remove useful air without solving the moving texture. If a click happens once at a known join, fix that join rather than processing the entire song. If energy below 30 Hz is only visible but does not trigger a limiter or weaken the audible bass, leave it alone.

ACE-Step documentation explains different inference paths, including turbo and base variants with different step ranges. It does not document a universal residual noise curve that proves a lower step count creates a predictable amount of hiss in every render. Treat settings as useful session notes, not as a diagnosis.

I usually begin with the busiest chorus because layered vocals, bright percussion, and long reverb tails make a synthetic edge easier to hear. Then I return to the intro at a lower monitor level. If the same texture remains distracting in both places, it is more likely to justify full-mix cleanup.

Use AI music artifacts explained if you need to separate hiss, metallic shimmer, clipping, and reverb smear by ear. Giving the problem the right name is more useful than choosing a processor first.

Diagnose audible problems in an ACE-Step render

Listen for repeatable symptoms, not a theoretical “ACE-Step sound.” The following checks work with headphones and an ordinary speaker; a spectrogram can confirm timing, but it cannot tell you what should stay.

• Moving high-frequency grain: Loop a vocal phrase, cymbal tail, or bright synth. A fine spray that follows the musical layer is different from steady room noise. Lower playback volume. If the grain disappears inside the arrangement and returns only when soloed, cleanup may do more harm than good.
• Narrow midrange sting: Some notes may jump forward with a hard, nasal edge while nearby notes remain comfortable. Sweep gently only to locate the area, then use a dynamic EQ that turns down the band when it becomes sharp. A static cut can make the entire vocal or snare dull.
• Low-end rumble: Check below about 30 Hz while watching whether your limiter reacts. Sub-audible energy matters when it consumes headroom, makes the bass unstable, or causes stereo-to-mono cancellation. A line on a spectrum display is not automatically a defect.
• Edit or repaint click: Mark the exact start and end of every changed region. A single click usually belongs to a short fade or crossfade. A larger change in pitch, room sound, rhythm, or arrangement belongs to regeneration or source editing.
• Softened attacks: Compare kick, snare, and plucked instruments in context. A gentle arrangement may have intentionally rounded transients. Expansion cannot recreate a clean attack that was never present, and aggressive settings can pull up noise between hits.

A useful rule is simple: if you cannot point to a timestamp and describe what becomes less distracting, do not process yet. A spectrogram pattern does not prove that ACE-Step caused it, and it does not prove that removing it will improve the song.

Source preflight: version, settings, format, and rights

Before cleanup, record the model or interface, seed, duration, inference steps, guidance settings, and any reference or source audio. These notes let you regenerate intelligently if the problem is musical. They also prevent you from attributing a later MP3 conversion or hard edit to the generator.

Preserve the original render. If your interface provides WAV, keep that file as the working source. If you only have MP3, do not convert it to WAV or FLAC and assume the discarded detail has returned. Work conservatively, keep the MP3 untouched, and export a new delivery file only after cleanup.

The official tutorial explains that reference audio controls broad acoustic features and that source audio can guide structure for cover tasks. That makes a preflight especially important: an unwanted room tone, over-bright reference, or contradictory prompt may be better fixed by changing the input and generating again. Cleanup should not become a substitute for a source decision.

ACE-Step 1.5 code is published under the MIT license, but software licensing is not the same as clearance for every prompt, lyric, source recording, voice, or output. Check the terms of the interface and assets you actually used. Keep authorship and license records with the project.

ACE-Step render inspection map

What you hear	What to check	Smallest useful action	Stop when
Sandy top end in vocals or cymbals	Same phrase in quiet and dense sections; original versus later encode	Gentle dynamic high-frequency control or a conservative cleanup pass	Air and consonants begin to disappear
One click at a repaint boundary	Exact sample region around the join	Short fade or equal-power crossfade	The join passes in context without a level dip
Unstable sub-bass or rumble	Limiter movement, mono playback, spectrum below 30 Hz	Cautious high-pass filter below the musical bass	Kick or bass loses weight
A few painful midrange notes	Loop only the notes that jump forward	Narrow dynamic EQ triggered by those moments	The vocal or instrument starts sounding recessed
Wrong note, rhythm, lyric, or arrangement	Original prompt, source audio, and generation settings	Regenerate or edit the source	Do not use cleanup to hide a musical decision

This map is intentionally based on what you can hear and verify. It is not an “Inference Step Residual Noise Curve,” because the official material does not provide evidence for a universal relationship between step count and audible noise across every model variant and song.

Restrained manual cleanup in your DAW

Duplicate the original track and work on the copy. Turn off mastering limiters, loudness maximizers, and broad saturation while you diagnose; those tools can exaggerate the very texture you are trying to judge.

For sub-rumble, start the high-pass filter below the lowest musical bass. Raise it slowly while switching between bypassed and processed audio. Stop as soon as kick weight or bass sustain changes. The purpose is to recover wasted headroom, not to make the spectrum look tidy.

For a narrow harsh area, a dynamic EQ lowers the frequency only when it becomes sharp. Use the smallest reduction that makes the repeated note easier to hear. The broader workflow in fixing harsh highs in AI music explains why multiple static cuts can remove brightness from the whole mix.

For a local click, zoom in and place a short fade or crossfade across the discontinuity. Listen before and after the join with the surrounding bar. If ambience or harmony changes, a longer crossfade may only smear two incompatible regions. Go back to repainting or regeneration.

Do not add upward expansion by default. It can restore a little contrast when attacks are present but overly compressed, yet it also raises quiet noise and room tails. Compare at equal loudness and stop if the groove becomes jumpy.

The Sunofix path for ACE-Step audio

I built Sunofix for the stage where the song already works but the exported full mix still has a synthetic edge that moves through several layers. Upload a lawful WAV or MP3, choose a conservative pass, and compare the same chorus, quiet passage, and edit boundary with the original.

Sunofix is most useful when broad manual EQ would remove too much music because the unwanted texture changes over time. It aims to reduce audible artifact texture while preserving melody, lyrics, arrangement, performance, and emotion. It does not replace a crossfade for one isolated click, stem mixing, or regeneration for a wrong musical event.

Level matching matters. A louder file can seem clearer and more exciting even when it is harsher. Compare the before and after at the same perceived level, then check headphones, an ordinary speaker, and mono. I stop when the distraction no longer pulls attention away from the song; chasing a perfectly empty spectrogram usually costs more musical detail than it saves.

Cleanup should come before final mastering. Artifact removal and mastering solve different jobs: cleanup reduces unwanted texture in the source, while mastering shapes the finished track for delivery. A limiter cannot restore detail already lost to clipping or lossy encoding.

Product, rights, and technical boundaries

Sunofix processes the mixed audio you provide. It cannot separate a stereo render into clean independent instruments, rewrite lyrics, repair timing, change chords, or restore a transient that is absent. It cannot recover detail already lost to clipping or lossy encoding.

A cleaned track is not evidence of ownership or permission. Sunofix does not grant legal clearance or guarantee distributor approval. It does not remove provenance marks, bypass detectors, or make an unauthorized source lawful. Confirm rights to prompts, lyrics, source recordings, voices, and reference audio under the terms that apply to your actual workflow.

There is no single correct noise floor. Breath, amp hiss, vinyl texture, distorted cymbals, and deep synth movement may be intentional. Keep them when they support the song. The goal is not sterile audio; it is a source that holds up under mastering without making the listener fight the texture.

ACE-Step release checklist

1. Preserve the original render and note the ACE-Step version, model path, seed, and inference settings.
2. Mark the quiet passage, busiest chorus, and every repaint or edit boundary.
3. Compare the same timestamp at matched loudness before judging any processor.
4. Name the symptom as moving grain, narrow harshness, rumble, a local click, or a musical error.
5. Regenerate musical problems and repair isolated joins locally before full-mix cleanup.
6. Filter sub-rumble only when it steals headroom or destabilizes the audible bass.
7. Use dynamic control sparingly and stop when air, weight, or attack begins to disappear.
8. Run a conservative Sunofix pass when the artifact moves through the full mix.
9. Check headphones, an ordinary speaker, and mono with mastering disabled.
10. Master only after the source decision is stable, then archive the original, cleaned source, and final master separately.

If the quiet section feels calmer, the busy chorus keeps its energy, and edit boundaries pass without drawing attention, you have a better source for mastering. If the result sounds smaller, duller, or less alive, return to the original and make a narrower change.