In February 2025, Dr. Yetta Wong and her colleagues published Learning fast and accurate absolute pitch judgment in adulthood in Psychonomic Bulletin & Review. The study trained twelve adult musicians over eight weeks on a structured online protocol, then tested whether any of them could reach absolute pitch performance at the speed and accuracy used to identify lifelong AP possessors in the published literature. Two of them did, on all twelve pitches.
The result itself is not surprising in isolation. Van Hedger, Heald, and Nusbaum (2019) and Wong, Lui, Yip, and Wong (2019) had already produced adults whose post-training pitch identification was behaviorally indistinguishable from lifelong possessors. Nevertheless, skeptics had a long list of objections to those initial results that could not all be ruled out by the existing data. The contribution of the 2025 study is that the protocol was built specifically to close those types of loopholes.
Skepticism About Adult AP Training
A handful of objections surface whenever a study reports an adult successfully learning AP, and the Wong 2025 design addresses the most common ones directly. It's worth noting that these measures have not been required for successful learning historically and were included in this protocol specifically to objectively measure the vectors for each objection, namely:
- Pitch height instead of chroma. If every note you train is in one octave, then naming notes is possible through memorization of specific frequencies instead of recognition of the invariant that all instances of the same note share. Training tones were drawn from three octaves spanning C4 to B6, and the post-test also evaluated tones from across all three octaves. A participant who had only memorized pitch height would have failed on the octaves they had not specifically memorized.
- Internal use of relative pitch. A participant could in principle remember the previous tone and compute intervals from it, even with no external reference. The training instructions discouraged this strategy and the post-test enforced it directly. The post-test separated successive tones by more than twelve semitones and capped the response window at five seconds. One participant (S3) attempted exactly this strategy and used relative pitch internally to "legally game the training system." He successfully completed all 288 training levels, but when the post-test removed the trial-by-trial structure that supported his strategy, his chroma recognition did not hold.
- Passing the criterion by luck. A one-time pass of a raw accuracy threshold could theoretically occur by a statistical fluke if attempts are unlimited. The protocol required passing the final 288th level four consecutive times with a mandatory wait of at least twelve hours between the third and fourth attempts. This is the direct rebuttal to "they got lucky." Four consecutive successful evaluations across a long no-training gap exhaust that explanation.
- Pre-existing AP ability. The study explicitly measured the starting AP performance of each participant. Group pre-test accuracy was 13.9%, barely above the 8.3% baseline of random guessing among twelve options. Average pre-test semitone error was 2.62 on the trained timbre, close to the chance baseline of guessing across the chromatic scale at random. Semitone error is arguably the more meaningful of the two measures, but by either measure, no participant entered the study with usable AP.
- Short-term auditory memory of reference tones. Training included optional sample tones the participant could listen to before any level. Whenever the next level was a no-feedback level, a 20-second Shepard tone was played afterward to disrupt the working-memory trace of those samples. The post-test went further. It allowed no feedback and no reference tones at all, and was conducted over Zoom with the participant's full desktop, environment, and face on camera throughout, so the session could be audited rather than trusted.
None of these measures are required for successful training. Zoom monitoring during the post-test, for example, is not part of how a learner internalizes pitch. The Shepard tone disrupts a specific test-design concern that is irrelevant to a private practice session. The four-times-in-a-row gate is a forensic safeguard for the published result, not a step in learning. The protocol layered them in so that the results, when reported, would effectively seal off all the standard escape hatches.
If the development of AP ability is not constrained by the critical period and remains learnable in adulthood, there should be no 'glass ceiling' for AP learning up to the performance level comparable to that of real-world AP possessors.
The Protocol
Forty-eight adults attempted the study and twelve completed enough training to be included in the analyses. The included group averaged 27.8 years old, ranged from 19 to 44, and reported an average of 12.8 years of music training and 5.08 hours of weekly practice. Seven were tonal language speakers (Cantonese) and five were non-tonal language speakers, with Hebrew, German, and English among their first languages.
Training was conducted entirely online over eight weeks, with a target of 25 hours total and at least 2 hours per week. The curriculum consisted of 288 training levels organized as 24 levels per pitch. The accuracy requirement at each level rose progressively from 20% up to 90%, and tones during training were piano samples from a Roland FP60 covering three octaves.
Training began with a single pitch. Only F was presented, alongside four "out-of-bound" pitches at one and two semitones above and below the target. The participant had to indicate either "F" or "out of bound" for each tone, learning to recognize chroma against tightly packed distractors. When F was mastered, E was added next, then F#, alternating between adding a pitch above and below the existing set. This cycle continued until all twelve pitches were included in the lessons.
Response times were strictly enforced. At the final, hardest level for any given pitch count, the response window was tight enough to make deliberate interval reasoning impractical, ranging from 1,183 ms when one pitch was being trained up to 2,028 ms when training all twelve pitches. Sample tones were available to listen to before the start of any level.
The pre and post-tests used different acoustic sources than training. Trained-timbre test trials used Yamaha Arius piano tones, a different piano than the Roland FP60 used during training, to confirm that the learning was general to the piano timbre rather than tied to a specific source. Untrained-timbre test trials added guitar tones from an online synthesizer to test generalization to a different timbre entirely. To be included in the analyses, a participant had to either complete 25 hours of training or pass the final 288th level under the aforementioned four-consecutive-attempts criterion.
What the Results Show
Across the included participants, the group spent an average of 21.4 hours in training (range 12.6 to 26.5 hours) and completed an average of 15,327 trials (range 7,500 to 24,524). On average, participants were exposed to 11.08 pitches during the curriculum and learned 7.08 of them at 90% accuracy or above within the assigned response time window, without trial-by-trial feedback or any external reference. The number of pitches learned ranged from 3 to 12 across the group.
At the group level, the pre-test to post-test improvement was substantial. Average size of error dropped from 2.62 to 1.50 semitones, a 42.7% decrease. Learning generalized partially to the untrained guitar timbre, with size of error dropping for both timbres and more so for the trained one. Semitone error is the most revealing measure because a learner whose errors cluster within a semitone of the target is in a meaningfully different place from one whose errors scatter across the octave, even when overall accuracy scores look the same.
S1, a non-tonal language speaker who did not start music training until age 24, reached a post-test accuracy of 89%, an average error of 0.42 semitones, and an average correct response time of 1,586 ms. S2, a Cantonese-speaking violinist, reached 78% accuracy, 0.31 semitone error, and 1,502 ms. Both performance profiles fall squarely inside the range that prior published research has used to identify lifelong AP possessors.
Training Is Possible
The authors summarize the central finding in two complementary statements:
Two participants passed all training levels and showed highly accurate and fast AP judgment with all 12 pitches in the post-test [...] demonstrating that there was no 'glass ceiling' for AP development in adulthood.
The two highest-achieving participants in the current study, based on the accuracy and speed of their AP judgment, would be categorized as 'true' AP in more than 80% of the published papers that adopted an objective performance-based definition of AP.
In other words, two participants achieved accuracy and response time comparable to lifelong AP possessors for all twelve pitches in eight weeks of online training. Notably, nothing in the data suggests that an upper bound was reached during training for any participant. S1 and S2 would be considered to have AP by the criteria that the majority of published AP studies have used to identify "true" AP possessors. The authors note that some researchers will likely dispute their interpretation, and they explicitly invite dissenters to articulate why a different threshold should apply to trained AP.
The continuous distribution of outcomes is its own finding. Participants learned between 3 and 12 pitches, demonstrating no bimodal split. Absolute pitch isn't a binary trait, and any spread at all in the post-test data is direct empirical evidence against treating it as one. Adult AP development under this protocol looks like a graded skill that responds to practice in different ways for different learners, not a switch that flips for a lucky few.
These results have been replicated across multiple unrelated peer-reviewed studies. Van Hedger et al. (2019) produced the same general finding through an independent protocol, and Wong et al. (2019) corroborated it across three separate experiments. Different labs, different participants, the same effect under controlled conditions. There is still more to learn, but the question of whether adults can learn absolute pitch is settled.