Singing the alphabet should be the easiest thing an AI voice can do. A-B-C-D-E-F-G. Seven discrete sounds, familiar to every child on the planet. Yet anyone who has experimented with AI music generation knows the reality is messier. Ask a model to sing the alphabet, and the letters L, M, N, O, and P will often collapse into a single blurry syllable. "Elemmennopee" is not a letter. It is a symptom.

This is the LMNOP problem, and it extends far beyond nursery rhymes. Days of the week, numbered instructions, and any kind of ordered list expose the same weakness. A developer recently put this to the test by generating eight songs through an AI pipeline and measuring the output with mlx-whisper, an automatic speech recognition tool. Instead of relying on subjective listening, they let the transcription software report exactly what the AI had sung. The results were stark. Enumeration trips up synthetic singers in ways that normal language simply does not.

Why Lists Break AI Voices

Spoken language carries context. If someone mumbles, "I'm taking the dog to the park," and you miss the word "dog," the sentence still makes sense. The surrounding words fill in the gaps. Lists do not offer that safety net. Each item stands alone. If the model slurs the difference between "B" and "D," the listener does not get partial meaning. They get noise.

In the alphabet song, the LMNOP cluster is the most obvious failure point. The voice model treats the sequence as a single phonetic blob rather than five separate letters. Without contextual clues to anchor each sound, the synthesizer guesses at the transitions between consonants. It usually guesses wrong. The same thing happens with days of the week or numbered steps. The model rushes through the sequence, compressing distinct items into an indistinct slurry.

Measuring the Damage

To study this objectively, the developer generated eight songs and ran them through mlx-whisper to score lyric accuracy. The pipeline was straightforward: write a prompt, generate the audio, transcribe the result, and compare the transcribed text against the intended lyrics.

For standard narrative lyrics, the output was reasonably faithful. Words landed where they should. But when the prompts included lists, alphabets, or enumerations, mlx-whisper returned gibberish. Letters vanished or merged. Numbers turned into unrecognizable syllables. The experiment confirmed that list-heavy lyrics suffer measurably worse intelligibility than prose.

Prompt Architecture That Helps

You cannot rely on post-processing to fix a garbled list. The fix has to happen before the first note is generated. A carefully constructed first prompt can force the model into clearer diction. These adjustments cost nothing, but they change how the model breathes and pauses.

  • Break long runs into smaller groups. Instead of A-B-C-D-E, structure the line as A-B-C-D and E-F-G. The tiny reset between groups gives the model a chance to land on the right consonants instead of smearing them together.

  • Spell out numbers. Use "thirty" instead of "30." Written digits are abstract symbols; the model sometimes compresses them into clipped, unvoiced noise. A full English word provides phonetic substance.

  • Insert visual spacing around letters. Write "A - B - C" with hyphens or spaces rather than "ABC." The visual separation signals to the model that these are standalone items, not a single acronym or word.

  • Lock the syllable count. Keep each line between six and ten syllables. Wild variation causes the model to rush short lines and stretch long ones. Lists already require precision; uneven rhythm makes accurate delivery almost impossible.

  • Strip the word "and" from lists. In natural speech, "and" acts as a conjunction. In a sung list, it adds a soft syllable that often swallows the hard edge of the previous item. "Monday, Tuesday, Wednesday" cuts cleaner than "Monday, Tuesday, and Wednesday."

The Regeneration Trap

Here is where human intuition backfires. With ordinary lyrics, iterative prompting usually improves results. You hear an off phrase, tweak the wording, and generate again. The next version sounds better. The test showed that this approach works for non-enumeration songs. But for songs containing lists, rewriting the prompt made the output harder to understand.

The data was unambiguous. Non-list songs improved after prompt fixes. Enumeration songs degraded.

The culprit is the acoustic seed. In lyrics-to-song models, changing the prompt does not simply edit the existing audio. It reshuffles the entire generative foundation. The model rebuilds the performance from scratch. For narrative text, the new dice roll might land on a clearer take. For lists, you are usually inviting a worse slur. You might fix the timing on one letter only to watch the model bury "J," "K," and "L" in the next attempt. The original take was flawed, but the revision often traded one phonetic mess for another.

A Smarter Workflow for List-Heavy Lyrics

Because regeneration risks destroying clarity, your process needs to change. Stop chasing the perfect rewrite. Instead, treat the first prompt like a casting call.

Generate multiple versions from the exact same prompt using different random seeds. Do not touch the text. Hold the lyrics constant and let the model vary its performance. You are running an audition, not an edit session.

Then score every candidate. Run each version through mlx-whisper or a similar transcription tool and measure which take matches your prompt most faithfully. Pick the winner based on lyric accuracy, even if the instrumentation or vocal tone is slightly less polished. For list-heavy content, intelligibility matters more than vibe.

Most importantly, front-load your fixes. Apply spacing rules, syllable limits, and grouping before you ever hit generate. Do not write the alphabet song in plain English and try to patch it later. The model is less forgiving of retroactive edits when lists are involved. A