Whistled languages

Why can a whistled message carry a full sentence across a valley without a single consonant in it?

A whistle is one thing at a time: a pitch that rises or falls, a loudness that swells or breaks. That is the entire signal. Speech is a broadband noise full of hisses, bursts and resonances, and we are taught that its meaning lives in consonants and vowels. Yet on La Gomera in the Canaries, in the Turkish village of Kuşköy, and among speakers of several Mexican and Southeast Asian languages, people conduct real business across a valley by whistling — unpredictable messages, not fixed calls. So what is crossing the valley, if the consonants did not?

Begin by asking what a whistle can carry, and be strict about it. Two channels only: frequency over time, and amplitude over time. No spectral shape, no voicing, no noise. Whatever whistled speech is, it must be a projection of the spoken signal onto those two channels. The interesting question is not whether that loses information — of course it does, enormously — but whether what survives happens to be the part the language was leaning on.

So where do a language's contrasts actually live? Not in letters. Vowels are distinguished chiefly by the two lowest resonances of the vocal tract, the formants F1 and F2, and for a small vowel system like Spanish the ordering of the vowels is largely recoverable from F2 alone: /i/ at the top, /u/ and /o/ at the bottom, /a/ between. Suppose a whistler moves the whistle's pitch up and down in step with that dimension. He has thrown away the vowel's timbre and kept its position in the vowel space. Would that distinguish pico from poco? It might, if little else in the sentence is compatible with the wrong choice.

Then the consonants — supposedly the missing half. Where does a listener's evidence for a consonant come from? Surprisingly little from the consonant's own noise; a great deal from the transitions, the way the formants bend into and out of the neighbouring vowel, and from timing — an abrupt break in amplitude for a stop, a gentler dip for a continuant. Those are precisely the two things a whistle keeps. So the consonants are not absent from the whistle. They have been relocated to the shape of the movement between vowel targets, which is close to where they were in the speech to begin with.

That gives the account a sharp prediction, and this is the part worth testing. If whistling transposes whichever cues a language happens to lean on, then a tone language should whistle differently from a non-tone language. And it does: in the whistled registers of Chinantec, Mazatec, Hmong and others, what the whistle carries is the tone melody, because that is where the lexical contrast sits, while the vowel information largely goes. Two languages, two entirely different whistles, each tracking its own language's load-bearing dimension. A whistled register is therefore not a code bolted on to a language; it is that language pushed through a two-channel pipe.

Why whistle rather than shout? Because of what distance does to a signal. A whistle puts a large amount of energy into a narrow band where the ear is sensitive; a shout spreads the same effort across the whole spectrum, and the high-frequency components — exactly the ones carrying consonant detail — are the first lost to attenuation and scattering. The whistle is already stripped down, so it has less left to lose. Whistled exchanges over distances on the order of kilometres are reported in favourable topography, and I am reporting that rather than deriving it.

Which explains the geography: whistled speech clusters in steep valleys, mountainsides and dense forest, where the walk between two points is long relative to the shout. Somewhere flat and open, the skill is not worth maintaining.

Three refinements.

First, a whistled language is not a language. It is a register of one: a Gomeran whistling Silbo is whistling Spanish, and a whistler of Chinantec cannot understand him. This is why the practice is productive — a whistler can transmit a stranger's name or an unexpected instruction, the sharpest test that it is a transposition and not a learned vocabulary of signals.

Second, "carrying a sentence without consonants" is the wrong frame; the honest version is that the whistle carries a heavily degraded version of everything, and the listener completes it. Experimental work on whistled Spanish and Turkish, chiefly by Julien Meyer and colleagues, finds isolated whistled words recognised well above chance but well short of perfectly, with accuracy rising sharply once context is supplied. Whistled exchanges are correspondingly short, repetitive and topic-bound.

Third, where I have simplified. Exactly which acoustic dimension non-tonal whistlers track is still argued — F2 is the standard account and fits well, but some analyses describe the whistler as tracking a more global perceptual quality of the vowel. And separating how much of the comprehension comes from the signal and how much from context is genuinely hard, which means my "redundancy" story could be carrying more weight than the evidence supports.

What would refute the account? Find a whistled register of a strongly tonal language that ignores tone and tracks vowel quality instead, or a non-tonal whistled register whose pitch contour bears no systematic relation to the vowels of the words being whistled. Either would show the whistle is an arbitrary code that merely happens to be whistled, rather than the language's own cue system projected onto pitch.

• Why whistled registers are dying faster than the languages they belong to.
• Drummed and horn-based speech surrogates, which keep timing and tone but discard pitch height differently again.

Every term the collection defines is gathered in the glossary.