Music theory training data. Notes are easy. Why they work is not.
A generation model can produce a chorale with parallel fifths in every phrase and never know. eeko commissions music theorists, composers, arrangers, and performers to label harmony and form, align notation to audio, write original works, and grade generated music, with rights recorded on every item and a second musician on every label.
What is music theory training data?
Music theory training data is music with the analysis attached. A score or recording where the key, every chord, every cadence, and every modulation is labeled. Part-writing marked for the techniques it uses and the errors it contains. Sections, phrases, and themes bounded across a whole piece. Notation aligned to a MIDI file or a live recording at the level of individual notes. And generated music graded by a musician who can say exactly what is wrong with bar seven.
Models trained on raw audio learn what music sounds like and nothing about why. The symptoms would get a first-year theory student marked down: chords named in the wrong key, cadences that never resolve, a meter the model has misread, MIDI that drifts out of step with the score halfway through, and generated pieces that wander without form. The audio usually comes from sources with disputed rights, too.
eeko's data practice addresses both problems at once: trained musicians label to a written scheme, a second musician checks every item, and every recording, score, and composition carries a consent and licence record from the creator who holds the rights. It is one field within our AI training data practice, sourced like our speech and audio datasets and licensed on the terms of our AI data licensing.
Music data for AI, from harmony to commission.
Six lines of work, scoped by style, instrumentation, and format, and proven on a pilot batch. Order one layer on music you hold the rights to, or commission music and analysis together.
What a music model gets wrong and a theorist marks in red.
Invisible to a loss curve, obvious to anyone who has taught a theory class. Our analysis scheme and grading rubric are organized around them.
Six errors the analysis scheme is built around
- Parallel fifths and octaves — the most basic part-writing error, produced confidently and at scale.
- The wrong key — chords labeled relative to a tonic the passage left two bars ago.
- Cadences that do not land — phrases that stop rather than resolve, because the model never learned what a cadence is for.
- Misread meter — a passage in three analyzed as if it were in four, with every downbeat displaced.
- Alignment drift — a score and a recording that agree at bar one and are a beat apart by bar forty.
- Music without form — a generated piece that wanders, never returns, and ends because the token budget ran out.
Paid remote work for musicians, between seasons.
The eeko expert network pays theorists, composers, arrangers, and performers to analyze, align, compose, and review. Rate, scope, deadline, and the licence terms for anything you compose are on every job before you accept it.
You do not need every item below. A theory-strong music degree or a professional portfolio is enough to start receiving matching jobs.
- Degree Music theory, composition, or performance
- Advanced study Graduate work or teaching experience in theory
- Professional work Composer, arranger, copyist, orchestrator, or transcriber
- Tools Fluent in notation software and MIDI
- Styles Common-practice, jazz, film, popular, or other specialisms
Written scheme or rubric on every job, a second musician on every item, and approved work credited to your account automatically.
- Analyze Keys, chords, cadences, voice leading, and form in scores and recordings
- Align Notation to MIDI and to performances, drift corrected
- Compose Original works or arrangements to a brief, with score, MIDI, and stems
- Grade and review Judge generated music, or check a colleague's analyses before delivery
Rights, fingerprints, and review on every recording.
Music carries more disputed rights than almost any other training material, and analysis is interpretive. Both need a record that satisfies a licensing lawyer and a second theorist.
From a repertoire brief to a signed delivery.
Fixed price per accepted item, agreed on the pilot and held for production. Volume starts when your team has reviewed real output against the scheme.
Music theory training data questions.
What is music theory training data?
Music theory training data is music with a musician's analysis attached: a score or recording with the key, chords, cadences, and modulations labeled, part-writing marked for its techniques and errors, and notation aligned to the audio note by note. Raw audio teaches a model what music sounds like; this teaches it why a passage works.
How much does music theory training data cost?
We price per accepted item, and the unit depends on the task. Labeling the harmony of a chorale phrase costs less than aligning a full orchestral score to a live recording, and an original composition is priced as a commission with its licence terms stated in the job. After scoping we quote a fixed price for a pilot batch, and the same unit price holds for production. Rejected items are not billed.
Who owns the rights to the music in the dataset?
Every recording, score, and composition is either licensed from the creator who holds the rights, with licence and consent attached to that item, or written new by a musician in the network under a job whose licence terms were stated before they accepted it. Audio is fingerprinted acoustically on intake and embedded metadata is read to confirm human origin. Nothing is scraped from streaming services or sheet music libraries.
What kinds of music data can you deliver?
Roman numeral and functional harmonic analysis, voice-leading annotation, score-to-MIDI and score-to-audio alignment, form and phrase labels, original compositions and arrangements with scores, MIDI, and stems, and graded comparisons of generated music with written rationales. Each request is scoped by style, instrumentation, and format, then staffed with musicians verified in that repertoire, and the pilot batch confirms the scheme before volume.
What qualifications do musicians need, and how are they paid?
Analysis jobs generally expect a theory-heavy music degree; for composition and arrangement a professional portfolio counts as much as any degree; alignment tasks need fluency with notation software and MIDI. Every job lists its rate, scope, and deadline before you accept it, there is no quota, and approved batches are credited to your account with payouts after identity and tax verification.
Can we license commissioned music and analysis exclusively?
Yes. Commissioned works, analyses, and alignment maps can be licensed to one buyer, with scope, term, and exclusivity written into the agreement before work starts. Audio ships as WebDataset or Parquet, labels and alignment maps as JSON Lines or CSV, and every delivery includes Croissant metadata and a signed receipt naming each item by identifier and hash.
Related expertise, solutions, and industries.
Send us the piece your model is proudest of.
Tell us the styles, label layers, and volume. A data lead replies inside one business day with a staffing view, a scheme outline, and a pilot price. Musicians: the application takes ten minutes. No deck. A ranked list.
Experienced within
Markets served.
As an enterprise AI agency, eeko systems delivers production AI systems remote-first across the United States and internationally — including these markets: