The earlier production log focused on putting a song brief into words. Later sessions raised a different question: once a promising version exists, what should the next attempt change, and how can I tell whether that change happened?
These are observations from this batch of work, not a general ranking of model capabilities. Settings from older models and interfaces also need checking before reuse.
Locate the difficult part first
I initially estimated difficulty from how unusual a genre was and how dense its arrangement looked. Neither was reliable. Some styles with clear conventions and one lead singer quickly produced usable takes; some dense arrangements did too. The slow work often involved keeping a genre’s identity while rejecting several of its most familiar sounds.
A symphonic metal brief might ask for force without a trailer-style male voice, repeated orchestral stops, or heroic harmony. Each request makes sense alone. Together, they require repeatedly pulling the result away from neighboring styles. Adding another instrument is not necessarily the harder task.
I now write three short lists:
- Keep: the melodic lead, rhythmic drive, and emotional direction.
- Allow to vary: ornament, local texture, and short transitions.
- Reject if present: strained high notes, the wrong vocal part, or a chorus that changes the intended direction.
These lists determine the first small test. If one vocal passage is uncertain, test that passage before spending more attempts on complete songs.
Give instructions different jobs
The earlier log placed a fairly long negative list in Style. Later attempts made me withdraw that as a general recommendation: some outputs emphasized features repeatedly named in the list. This observation does not establish how the model processes negation internally.
| Location | Main job | Approach used in this batch |
|---|---|---|
| Style | Overall genre, lead, timbre, and momentum | Audible positive descriptions, such as a snare on beats two and four |
| Section labels | Who performs and what drives this passage | One or two arrangement cues rather than a miniature score |
| Exclude | Unwanted instruments or styles | Keep a separate record alongside Style |
Suno’s documentation describes the Exclude input. Having that input does not remove the need to listen to the result.
An instrument meant only for the ending should not also be heavily emphasized as a defining feature of the whole song. Local requests spread into other passages in this batch. Moving a request to its section and checking what remained intact was easier to evaluate than adding more restrictions.
Ask one question per iteration
Once a baseline works, change one item: the drum emphasis, a lead instrument, or the density of one section. Hold other settings steady where possible. Generation remains variable, so one better take does not prove causation. Changing five things together, however, makes the failure difficult to diagnose at all.
Listen twice: first to the target passage, then to the complete song. A local improvement that damages continuity should not automatically replace the baseline. After several drifting attempts, return to the last version worth keeping.
Save the model and mode, original Style, lyrics and labels, exclusions, reference-audio or Persona state, actual slider values, output ID, and selection reason. Actual values matter: the notes contained cases where a planned setting was never applied or an option intended to be off remained on.
At the time of this review, the official v6 FAQ says Variety adjusts style prompts and can be set to zero for full control of style tags. Old Weirdness and Style Influence values therefore should not become a universal recipe for the current interface. Check the controls available in the chosen mode first.
The description on a finished song page is also a poor replacement for the original input. Some page descriptions in the project did not match the audible result. That makes them unsuitable as primary experiment records; it does not reveal the entire generation pipeline.
A modulation instruction needs an audio check
Starting in a requested key, briefly pointing toward a new one, and sustaining that new center through a chorus are different outcomes. The notes include changes that never happened, changes that reverted a few phrases later, and individual successful transitions. Labels and page descriptions cannot distinguish them reliably.
Two local tools helped inspect the audio. This article explains their methods; it does not provide an online detector.
keyscan decodes mono audio at 22,050 Hz, uses an 8,192-point FFT with a 2,048-sample hop, estimates twelve pitch-class energies in the 65–2,100 Hz range, and corrects for tuning offset. It compares the result with 24 major and minor templates, both across the full track and in moving windows.
It asks which keys a passage resembles. Its leading and runner-up correlations are similarities, not probabilities of a correct answer. Relative keys, prominent arrangement tones, and short windows can complicate the result. One overall winner can also hide a change halfway through a song.
keytrack compares two candidate keys by the proportion of energy inside each seven-note scale. Its default window is six seconds with a three-second step. The current implementation labels a preference when the fit difference exceeds 0.03 in either direction; otherwise it leaves an ambiguous region. That threshold is a working heuristic, not a significance test.
The comparison can indicate a change in pitch material without establishing a new tonic. Relative major and minor share a pitch collection and cannot be separated this way. Borrowed chords and chromatic notes can also shift the fit. A different collection scoring better is not, by itself, proof of modulation.
Bring the measurement back to listening
Mark the intended transition, compare the windows on either side, then listen to the transition, melodic landing notes, and subsequent phrases. Does the new center last, or was there only a passing harmonic color?
When measurement and listening disagree, leave the result unresolved and, if necessary, compare pitches on a keyboard or in a DAW. No new accuracy benchmark was run for this article, and individual successes do not establish guaranteed key control.
The useful change was giving each retry a clear question: what changed, what became audible, and what evidence is still missing? Keep successful versions and failed judgments alike, so the next attempt starts from a record rather than an impression.