The Sentence Didn't Learn to Sing

“Sometimes behave so strangely.”

Those four words are the working material of a very economical musical illusion. In Diana Deutsch's demonstration, they begin as part of an ordinary spoken sentence. A recording of the phrase is then repeated. For listeners who experience the effect, what sounded like speaking comes to sound like singing. Nobody has added a guitar, corrected the pitches or asked the speaker to take another pass. The excerpt is the same. [1]

I like the economy of that setup. Usually, to compare speech with song, we change the sound: a person says something, then sings it. Here the interesting change happens on the receiving end. But the temptation is to let the illusion explain too much—to turn a striking demonstration into “anything becomes music if you loop it.” The experiments are more particular, and more interesting, than that.

Same ending, different route

In their 2011 study, Deutsch, Trevor Henthorn and Rachael Lapidis tested 54 listeners with at least five years of musical training. Each heard a complete sentence and then ten presentations of its seven-syllable fragment, “sometimes behave so strangely.” After each presentation, listeners rated it on a five-point scale running from exactly like speech to exactly like singing. [1, Experiment I]

The important control was not just that one group heard a loop. It was that all three groups heard exactly the same first and last versions of the fragment. What differed was the material between them.

One group heard exact repeats. A second heard intervening versions shifted slightly up or down in pitch, preserving the relative pitch intervals and the phrase's timing. A third heard its original syllables in changing jumbled orders. Exact repetition moved the ratings strongly toward song. The transposed versions left the final target in the speech region, though a little closer to song; the jumbled versions also left it sounding spoken.

The final recording had not changed. Its history had.

That makes the result harder to dismiss as a property you could simply read off the last sound file. It also puts a limit on a tempting explanation: the loop cannot work merely by supplying repeated timing, because the pitch-shifted versions kept that timing without producing the strong transformation. Nor was preserving the relative melody enough in this particular experiment. The listener's encounter with a stable, ordered pattern mattered.

“This particular experiment” is doing real work there. One phrase and a musically trained sample do not establish a law that every repeated utterance must obey.

Not just a different label

A second experiment asked people to repeat the phrase back as they had heard it. After one presentation, their renditions sounded spoken; after ten, they sounded sung to another group of listeners. Pitch analysis also found that the ten-presentation renditions were more consistent across people and closer to the original phrase's pitches. Their intervals were closer still to a proposed simple melody than to the original speech. [1, Experiment II and Table I]

There is something revealing in that combination. Repetition was associated with more faithful reproduction of some pitch information, but also with a tendency to regularize the intervals toward a melody. Hearing the musical shape need not mean making a photographic copy of the voice.

This evidence has a catch worth keeping beside it: the eleven people in the repeated-speech production group had already reported experiencing the illusion. Their performances help characterize what the effect can involve; they do not tell us how many people will experience it. And singing something back involves memory and vocal control as well as hearing. A reproduction is evidence, not a microphone placed inside perception.

The loop need not be a metronome

An independent study by Elizabeth Margulis, Rhimmon Simchy-Gross and Justin Black tested a different possibility: perhaps regularly recurring excerpts create the larger rhythmic structure that makes speech seem musical.

Their 2015 experiment used 24 young native English speakers, most without formal musical training, and recordings of the same north-wind-and-sun story in seven languages. Short excerpts were repeated ten times. One group heard regularly spaced repetitions; the other heard repetitions with conspicuously varying gaps. The timing inside each excerpt stayed intact. Listeners rated the short segment before repetition, then rated it again within a replay of the full utterance afterward—a different protocol from the original study's ratings after each repeat. [2, Materials and Methods]

The study detected no overall effect of that spacing manipulation on the change in songlikeness. An evenly timed outer loop was not required for the shifts they observed. That is not proof that rhythm never matters: irregular gaps between excerpts are not the same thing as scrambling the rhythm inside a phrase, and a small experiment cannot establish that two conditions are equivalent in every respect. It does make “the loop adds a beat” a poor complete explanation. [2, Results]

The stranger result concerned pronunciation. The researchers initially expected languages easier for English speakers to pronounce to transform more readily: perhaps being able to imaginatively join in would help speech become song.

The result went the other way. Excerpts in languages categorized as harder to pronounce showed larger shifts toward song. The investigators' own plausible explanation had made the wrong prediction.

Before turning that reversal into another grand theory, look at the scale. For the researchers' initially hypothesized harder category, the mean rating rose from 1.56 to 2.40, where 1 meant exactly speech and 5 exactly singing. That is a movement toward song, but the final mean remains below the midpoint of 3. “More songlike” and “now unmistakably singing” are different claims. [2, Table 1]

There were other limits. These were selected excerpts, not many speakers and sentences representing each language. Pronunciation difficulty was relative to these English-speaking listeners, not an intrinsic musical property of Hindi, Irish or any other language. The researchers checked two acoustic measures, but did not experimentally isolate pronounceability from every other difference among recordings.

A change in what gets to matter

Both papers suggest that repetition can bring pitch relationships forward—features already present in the voice but not necessarily prominent during ordinary speech listening. They also propose accounts in which the usual processing of speech loosens its grip. Those are explanatory possibilities, not neural switches measured by these two behavioral experiments. The later study did not show that meaning disappeared or that a language circuit shut down.

The useful distinction is smaller and sturdier. A recording has acoustic features. A listener has a history with it. “Speech” or “song” is a perceptual judgment arising from that encounter, not simply a label permanently embedded in the waveform.

I came to this question expecting repetition to do something like adding an accompaniment. I leave it interested in what repetition lets a listener foreground. No new note needs to arrive for an existing pitch shape to take on a different role.

The next question is not whether enough looping can make everything musical. It is why one phrase yields its melody while another keeps insisting on being a sentence.

Sources and listening note

[1] Diana Deutsch, Trevor Henthorn and Rachael Lapidis, “Illusory transformation from speech to song,” Journal of the Acoustical Society of America 129 (2011), 2245–2252. Complete author-hosted paper. The author's demonstration page provides sound examples; it is not a separate replication.

[2] Elizabeth H. Margulis, Rhimmon Simchy-Gross and Justin L. Black, “Pronunciation difficulty, temporal regularity, and the speech-to-song illusion,” Frontiers in Psychology 6:48 (2015). Complete article.

This essay is based on reading the complete papers, including methods, tables and textual figure captions—not on my own playback, singing or listener experiment. The two studies are distinct experiments; papers they cite were not thereby separately examined here.