Human preference data · Mexico & Latin America
Your model scored 4.2.
Which 4.2?
An average is the one number that hides the thing you needed to know. Gustoo recruits, calibrates and runs expert and listener panels across Latin America, and returns the full distribution — by country, genre and segment.
Why regional evaluation
Spanish is not a segment.
Most evaluation sets were rated in English and Mandarin. Translating the rubric does not translate the judgement — and "Latin America" collapses six music markets that disagree with each other constantly.
Structure is genre-bound
A corrido, a bachata and a reguetón track each resolve coherently inside their own conventions. Rated against Anglo pop form, all three read as structurally broken.
Hooks are local
Memorability rides on cadence, slang and rhythmic phrasing. A hook that lands in Mexico City can pass unnoticed in Buenos Aires — the same audio, a different result.
Vocal naturalness is prosody
Breath, diction, accent and the fit between lyric and melody. Spanish prosody has its own failure modes, and raters who don't speak the variety won't hear them.
Baseline rubric
Five dimensions, calibrated regionally.
We start from the dimensions used in current song-evaluation research, then re-anchor every definition with regional examples before a single production task is assigned. Custom rubrics replace these when your decision needs something else.
Musical and emotional continuity across intro, verses, choruses, bridge and close — transitions, dynamics, unified tone.
Whether melodies, rhythmic motifs, hooks or vocal phrases stay with the listener after one pass.
Breath, phrasing, pronunciation, prosody, and the relationship between the voice and the lyric.
How legibly sections are organised. Experimental form scores high when its logic is musically comprehensible.
Listening satisfaction across melody, harmony, instrumentation, and how voice and accompaniment integrate.
How a program runs
Managed end to end.
You send outputs. We handle recruitment, calibration, production, quality control and delivery. Six stages, in order — nothing enters production before it passes calibration.
Design
Define the decision the data serves, the segments, rubric, language, confidentiality terms and acceptance criteria.
Recruit
Select experts, target listeners, or both. Verify identity, experience and cultural eligibility for the segment.
Calibrate
Anchor examples, training sessions and comprehension tests. Raters who don't calibrate don't enter production.
Produce
Assign tracks, capture scores, pairwise preference, written critique and time-on-task per rater.
Quality control
Agreement measurement, outlier review, hidden gold tasks and duplicates, expert adjudication of disputes.
Deliver
Dataset, rubric documentation, per-segment breakdown, agreement metrics and a walkthrough of what the sample showed.
Panels
Two kinds of judgement.
Experts — producers, composers, audio engineers, session musicians, vocal coaches and critics. They apply the technical rubric and can name the specific failure.
Target listeners — real consumers, segmented by country, age and genre affinity. They answer a different question: whether anyone would actually want this.
They are not substitutes. Improving vocal technique needs experts; predicting market preference needs listeners; most programs need both, weighted deliberately.
Recruiting reach
Delivery formats
Start here
A pilot is 300 tracks and four weeks.
Send a sample of outputs. We design the panel, calibrate it, run it, and return a structured dataset with the per-segment breakdown and the disagreements. If the data doesn't change a decision you were about to make, it wasn't worth running.
Prefer email? Write to e.munoz@gustoo.ai and ask for a sample dataset.