Prompting Songs in MusicStudio and Cleaning Them in SoundStudio

A music model gives you the most probable song for your words. Vague words get the most probable song in the genre, which is the song everyone else got too. Specific words narrow the target. That is nearly the whole craft of prompting, and the rest is knowing which tool to grab when the result is 90 percent right.

Anatomy of a style prompt

Part	Weak	Strong
Genre and era	folk	1970s country-folk, Bakersfield touch
Instruments (2 to 4 leads)	guitar	fingerpicked acoustic, pedal steel, upright bass
Rhythm and tempo	medium	shuffle, 96 BPM
Key or mode (optional)		G major
Mood and arc	happy	warm, wry, builds a little into the last chorus
Vocal	singer	weathered male baritone, close, dry
Production		live room, minimal reverb, no autotune
Exclusions		no synths, no electronic drums

Mureka's own FAQ suggests a simple starting set of style or genre, mood, and a scene or use case, and says the model is especially strong at blending styles. Voice Studio's MusicGeneration tips add the emotional arc and the intended use. "For a podcast intro" is a surprisingly strong instruction, because it implies a short, steady piece with a clear start.

Exclusions need care. MusicStudio's Style Builder lets you click a chip once to include it and again to exclude it, and the page explains what happens next: Mureka has no negative-prompt parameter, so exclusions are written into the prompt as plain instructions such as "no drums, no synth." That steers hard but is not an absolute filter, and excluding a whole family (all synths) works better than one narrow instrument. Suno handles it differently, with a separate Exclude field under Advanced Options. Either way, expect the occasional stray synth pad.

Lyrics the model can follow

MusicStudio's Song box takes up to 5,000 characters, with buttons that insert [Intro], [Verse], [Pre-Chorus], [Chorus], [Bridge], [Break] and [Outro]. Lyrics written by a person matter for ownership, as lesson 8 explains. Here is an invented example in the shape a model handles well:

[Verse]
The mill shut down in ninety-one
We kept the whistle anyway
It blows at noon for no one now
And no one asks it not to

[Chorus]
Ring it out on Mercer Street
Let the old brick hear it
We were here, we were here
Ring it out on Mercer Street

[Outro]
We were here

Some habits that help:

• Short lines. Six to nine syllables sit on a melody more easily than fifteen.
• The hook repeated exactly. The model recognizes a chorus partly by repetition.
• Numbers as words. "Ninety-one," as in narration.
• Stage directions in the style prompt. Words in the lyrics box tend to get sung.

The page also offers "Write with AI," which uses Mureka's lyric writer and returns a title with the words, and "Extend lyrics." They are fine for drafts. AI-written lyrics weaken any claim to authorship of the song, which is why many songwriters treat them as scaffolding to replace.

MusicStudio's engine settings

Setting	Choices shown on the page	Guidance
Mode	Lyrics to Song; Prompt to Song (AI writes lyrics); Instrumental (no vocals)	Instrumental for beds; Lyrics to Song for anything with words you wrote
Model	Auto; mureka-9.5 (newest, best fidelity); mureka-9; mureka-8 (needed for head or tail extend); mureka-o2 (reasoning, no voice or melody input); mureka-7.6 (legacy, widest feature support)	9.5 for finals; 8 if you plan to extend
Variations	1, 2 or 3 songs, billed per song	2 is the sweet spot: a real choice for twice the cost
Vocal gender	Auto, male, female	Set it; Mureka's FAQ says an audio reference overrides it if both are given
Conditioning	Reference, voice, melody, instrumental reference	Melody cannot be combined with a reference or a voice

The style prompt box holds 1,024 characters and rebuilds itself from your chip selections until you type in it, which locks your wording. Presets save every chip state plus tempo, key, time signature, engine settings and notes, so a series can keep one sound.

The Studio tab: fix instead of regenerate

Tool	What the page says	When to use it
Extend	Grow a song forward from the end or backward from the start; head extend needs mureka-8	The bed is 48 seconds and you need 60
Remix	Re-record the same song in a different style; instrumental sources not supported	Same song, acoustic instead of full band
Region edit	Rewrite one stretch, at least 3 seconds, and leave the rest alone	One bad line in verse two
Generate a single track	Add one part, such as vocals, bass, drums or brass, to an existing recording	Add a bass line under your own guitar demo
Split into stems	v1: 5 stems, WAV only; v2: up to 12 stems plus MIDI; v3: vocal plus accompaniment, WAV plus MIDI; results as a ZIP	Lower the vocal under narration; fix a click on one part
Analyze	Describe instruments and tags; recognize lyrics and timings; transcribe to sheet music (MusicXML and PDF)	Checking what you got, or getting a lead sheet for a live band

The Voices and References tab can also clone a singing voice from 15 to 30 seconds of clean solo singing, with the page's own reminder to use voices you have the right to use. That means your own voice, or a singer who agreed in writing. Lesson 7 covers why.

Two housekeeping notes from the page. Songs stay addressable by ID for about a month, and the Library warns that Mureka's audio links expire about 30 days after generation, so download anything worth keeping. And the Console tab logs every request with the reason for any failure, which tells you whether the problem was your input or the API.

Cleaning up in SoundStudio

SoundStudio's theory page has a section for exactly this job, 15. Cleaning up an AI-generated song. Here is its workflow, with reasons:

1. Get the best file. Download WAV if the service offers it. The theory page notes that Suno downloads are often MP3 and advises exporting your edit as 24-bit WAV and making an MP3 only at the very end, because re-encoding stacks damage.
2. Listen first, all the way through on headphones, noting times of anything that bothers you.
3. Clicks: Noise tab, Find Clicks with the detector set to Suno ticks, then again with Spikes above the outline. Review the strongest first, use Fix This Click on the real ones, and the Spike Pen for ones you can see that the detector missed.
4. Hard spots: if a click is buried in a busy chorus, separate stems in the Helper tab, fix it on the vocal stem where it stands out, then use Mix Files Together. The help page says stems added back at 0 dB rebuild the original mix.
5. Timing: shorten an overlong held note with Smooth Cut; remove a bar or repeat with a plain delete at a beat, crossfaded.
6. Tone: only if needed. A gentle low cut at 30 Hz and small EQ moves, since AI songs usually arrive already mixed.
7. Loudness: Enhance tab, target -14 LUFS, -1 dBTP ceiling, and read the report. Export a 24-bit WAV master, then MP3 320 or FLAC to share.

Why -14 LUFS? Spotify says it normalizes tracks to -14 LUFS by default and recommends keeping true peak below -1 dBTP. The theory page makes the practical point: a song squashed to -8 LUFS just gets turned down by 6 dB and sounds flatter than one mastered at -14 with its punch intact.