How we make the tests — and why we don't use official ones

Exam boards publish the format of their aural tests in their syllabuses: which tests each grade has, the range of the melodies, the keys, how many times the examiner plays, which features may be asked about. They sell the actual specimen tests in books. We built our practice from the published format only — every exercise you hear is created by our engine, and none is copied or adapted from a board's book.

1. Reading the formats

For each board we read the current syllabus documents directly (for ABRSM, the Practical Grades specifications for 2025–2026 and 2027–2028; for RCM, the 2022 Piano Syllabus and 2024 Musicianship Addendum; for AMEB, the 2026 Manual of Syllabuses; for Trinity, the Aural Parameters from 2017). Each grade becomes a configuration: tests, keys, ranges, metres, cadence types, modulation targets, number of playings, what is given first.

2. Generating music that sounds musical

• Melodies are composed over an implied harmony, from rhythm patterns found in real tunes: steps rather than random leaps, a leap followed by a step back, a single high point, non-chord notes only as passing or neighbour notes, cadences at the ends of phrases, and longer melodies built as question-and-answer periods.
• Harmony for cadences and modulations is voiced in four parts and checked against classical voice-leading rules: no parallel fifths or octaves, the leading note and chordal sevenths resolve, chords are complete and properly doubled. Modulations go through a pivot chord and are confirmed by a cadence in the new key.
• Pieces for the musical-features questions are either composed to order (so every feature — dynamics, articulation, tempo changes, tonality, character, texture, structure — is a known setting) or taken from public-domain repertoire (Bach to Debussy) with our own performance markings and answer keys.

Every exercise has a computed answer key. Automated tests generate ten thousand random exercises of every type and check that the key agrees with the music actually played — for example, that a crescendo really gets louder, that the changed note in a “spot the change” test is where the answer says, and that a perfect cadence really ends V–I.

3. Listening on your device

Singing is analysed with a pitch tracker (the McLeod Pitch Method) running in your browser: notes are separated, compared with the phrase without regard to octave, and a small overall tuning drift is forgiven. Clapping is detected as short bright bursts and compared with the rhythm after allowing for your own tempo. We chose the pitch tracker by measurement: on 260 test recordings it agreed with the expected verdict 94% of the time against 85% for the alternative (YIN), and it never passed a wrongly sung phrase.

4. Marking spoken answers

Your answer to a musical-features question is matched against the piece's answer key — in plain English (“it got louder”) or Italian terms (“crescendo”), with negations and order (“quiet at first, then loud”) understood. When the wording is unusual, an AI examiner (Anthropic's Claude) marks it against the same key; it receives only the text of the answer, never audio or anything identifying.

The full measurements are published in our quality report: agreement on singing, clapping and several hundred spoken-style answers written in the voices of children and adults, including partly right and wrong ones.

Quality, in numbers

• Singing (260 recordings: children's, women's and men's voices, wrong notes, out of tune, missing and extra notes, octave displacement): 94% agreement with the expected verdict — and no wrong phrase marked as right.
• Clapping (260 recordings: clap-back rhythms and clapping the pulse with the piano playing): 98.5% agreement.
• Pitch tracking on real recorded voices (Dagstuhl ChoirSet, 48,000 annotated frames): 95.7% of frames within a quarter-tone.
• Spoken answers about musical features (204 answers held back from development, in the styles of children and adults, including partly right and wrong ones): 90% exact agreement with the expected mark from the rules alone, 97% with the AI examiner checking unusual wording; every answer within one mark.

Mock marks are estimates. Examiners mark the aural section holistically and their judgement is final.