Mumbling as Data Compression.

Julie Sedivy has an interesting post at Nautilus:

Far from being a symptom of linguistic indifference or moral decay, dropping or reducing sounds displays an underlying logic similar to the data-compression schemes that are used to create MP3s and JPEGs. These algorithms trim down the space needed to digitally store sounds and images by throwing out information that is redundant or doesn’t add much to our perceptual experience—for example, tossing out data at sound frequencies we can’t hear, or not bothering to encode slight gradations of color that are hard to see. The idea is to keep only the information that has the greatest impact.

Mumbling—or phonetic reduction, as language scientists prefer to call it—appears to follow a similar strategy. Not all words are equally likely to be reduced. In speech, you’re more likely to reduce common words like fine than uncommon words like tine. You’re also more likely to reduce words if they’re predictable in the context, so that the word fine would be pronounced less distinctly in a sentence like “You’re going to be just fine” than “The last word in this sentence is fine.” This suggests that speakers, at a purely unconscious level, strategically preserve information when it’s needed, but often leave it out when it doesn’t offer much communicative payoff. Speaking is an effortful, cognitively expensive activity, and by streamlining where they can, speakers may ultimately produce better-designed, more fluent sentences. […]

The notion of strategic laziness, in which effort and informational value are judiciously balanced against each other, scales up beyond individual speakers to entire languages, helping to explain why they have certain properties. For example, it offers some insight into why languages tolerate massive amounts of ambiguity in their vocabularies: Speakers can recycle easy-to-pronounce words and phrases to take on multiple meanings, in situations where listeners can easily recover the speaker’s intent. It has also been invoked to explain the fact that across languages, the most common words tend to be short, carrying minimal amounts of phonetic information, and to account for why languages adopt certain word orders.

There are links to papers backing up various points mentioned, and a nice zinger at the end.


  1. There’s an interesting ‘dual’ version of data compression that’s become fashionable lately in signal processing– ‘compressive sensing’. In compressive sensing, a sensor detects only a compressed version of a signal rather than the full uncompressed version. This simplifies the work a sensor has to do by a large factor, since you can search a much lower dimensional space for the information that you are looking for. In the linguistic context, this raises the possibility that both sides of a conversation, both the speaker and the listener, are compressing– leading to the possibility that the full uncompressed signal never actually appears in the back-and-forth information transfer.

    For the record, here’s a link to an incomprehensible Wikiipedia article on compressive sensing:

  2. and a nice zinger at the end

    Reminds me of the first time I visited the southern states, during March break several decades ago. Whenever I ordered ‘tea’ in a restaurant, the waitress invariably asked whether I wanted ‘hot tea.’

    Not to be confused with Red Zinger tea, introduced about the same time.

  3. Matt,
    “There’s an interesting ‘dual’ version of data compression that’s become fashionable lately in signal processing– ‘compressive sensing’. In compressive sensing, a sensor detects only a compressed version of a signal rather than the full uncompressed version. ”

    The same thing happens in language, when a speaker of one language cannot distinguish the sounds of another because his language does not recognize that distinction.

  4. Exhibit A, the french language.

  5. “Reminds me of the first time I visited the southern states, during March break several decades ago. Whenever I ordered ‘tea’ in a restaurant, the waitress invariably asked whether I wanted ‘hot tea.’”

    Huh? I am a native speaker SoAmE and I am sorry, but I don’t understand.

  6. Some languages make do with just three or four distinct words for color; for example, the Lele language, spoken by tens of thousands of people in Chad, uses a single word to encompass yellow, green, and blue. Languages with minimalist color vocabularies tend to be spoken in pre-industrial societies, where there are very few manufactured objects to which color has been artificially applied.

    Is this correct? I thought that the difference was that minimal color word societies tend to be hunter-gatherer, while complex agricultural societies develop wider color vocabularies. Putting the difference at industrialization means that most color-specific words are relatively recent coinages.

  7. Paul Ogden: Okay, sorry I responded before reading the article. Yeah, ice tea is understood as the drink one would get. The only question nowadays is “sweet or unsweet.”

  8. J. W. Brewer says:

    Yeah, what fisheyed said. Or, at least, if you look at the list of eleven “basic” color terms found in English at, it’s been stable since maybe the 14th century (“orange” was I believe the last to arrive). Industrialization gave us e.g. “magenta,” but that’s not what the usual debate is about.

  9. George Grady says:

    The OED only has “pink” as the color going back to 1669, named after the flower, whereas citations with “orange” as the color go back to 1557. I didn’t expect that.

  10. David Eddyshaw says:

    “Hunter-gatherer” is overstating things. Lots of West African languages have three basic colour terms, and they’re mostly agricultural societies, not hunter-gatherers. Some of them have centuries-old complex near-feudal setups with longstanding military aristocracies, to say nothing of histories of big land empires and widespread literacy in Arabic.

  11. @MattF: The Wikipedia article is very detailed and technical, but you only need to understand this part: “In compressed sensing, one adds the constraint of sparsity, allowing only solutions which have a small number of nonzero coefficients.”

    In short, a practical application of Occam’s Razor — given an ambiguous signal, find the simplest explanation, for some useful definition of simple. The rest of the article is about giving the idea a mathematical foundation and applying it to signal processing.

    Historical linguistics is an application of compressed sensing in a wider sense (no mathematical scaffolding), and the difference between any two versions of PIE (for instance) is a consequence of a) how lossy an input was used b) disagreement on how to define simple.

  12. It bothers me that any Wikipedia article having to do with mathematics is essentially useless to anyone but a mathematician. An article in any other specialized field will usually start off with a general sort of explanation a layman can grasp, followed by the jargon-filled details for those who can use them, but mathematicians (or at least those who take up the mantle of Wikipedia editor) seem to be contemptuous of anything a layman might be able to understand and ruthlessly edit for maximum jargonicity.

  13. LH: I agree, but it’s very hard to be both correct and clear in writing mathematical text– and in a large projects like Wikipedia you will settle, generally, for correct. That said, it’s also fair to say that mathematicians are averse to explaining what they’re actually talking about.

  14. Many years ago now, I tried rewriting a few articles about mathematical topics to make them both accessible to laymen and mathematically correct. I was unable to do so. The simple difficulty of the undertaking was part of it, but the biggest problem was the other Wikipedia editors. People kept inserting inaccurate examples and explanations, based on their own misunderstandings of the topics. Years later, once these kinds of articles became impenetrable to people without certain graduate-level backgrounds, these kinds of intrusions seemed to have been driven off. (Or the added inaccuracies became much easier for the people maintaining the articles to find and prune away.)

  15. The thing is that it’s impossible to leap directly from ignorance to graduate-level knowledge without passing through several levels of “close enough for government work” approximations. The introductory paragraph of any article should give a general sense of what it’s about, with whatever inevitable simplifications and distortions involved and “see section X below for clarification” where necessary. The overly precision-conscious editors involved are pulling up the ladder behind them. “Sorry, if you want to get even the faintest idea of what we’re talking about you’ll have to get a PhD at MIT.”

  16. Stefan Holm says:

    In the aftermath ow WWII it was found desirable that ’the two worlds’, science and humanities, must come together. Those physicists and mathematicians who had constructed so many deadly weapons and finally the nuclear bomb must be introduced to the world of values, beauty, meaning of life etc. And on the other hand the humanists and lawyers must get a glimpse of the scientific world.

    So in 1967 I in my home town belonged to the first cohort studying science at upper secondary school that was obliged to take classes in philosophy, litterature, psychology, religion and history of art and music. My very good friends among the ‘humanists’ got the corresponding basic orientation in science and mathematics.

    Thanks God, I say. Without this bridging idea I maybe wouldn’t have been able to appreciate (and participate in) this eminent blog.. It however turned out to be harder the other way around. My dear long time friends just had a harder time to grasp the world of mathematics. May I to illustrate this bother you with three quiz questions I’ve frequently used at parties?

    1) In Sweden there is a rift lake called Vättern. It’s about 130 kilometers long. Imagine that Peter at the north end and Mary at the south end decided to stretch a rope between themselves. (I know it’s physically impossible but you could send a laser beam all the way). Anyway, since the earth is round (you believe in that?) – the rope would be under the water surface on its way (the rope is straight, the earth is round). How far under the water surface would the rope be at the middle of lake Vättern?

    2) The circumference of the earth is approximately 40,000 kilometers. That is the length of a rope you would need to stretch all the way around the globe (forget about mountains and valleys – it actually doesn’t change the quiz). Now say that you want to stretch the rope one meter above the surface of the earth. How much additional rope would you have to buy?

    3) If you flip a coin the chance is 50-50 that you would get a head or a tail. Now, if you flip a thousand coins – what is the probability that 400 or less (40%) would turn up as tails?

    You out there who know about geometry and probabilites – hold your horses and let us see what our human gut feeling tells us. I for one don’t agree with the inscription above the entrance of Plato’s Akademia: MÉDEIS AGEOMÉTRETOS EISITO – non-mathematicians are not allowed. But I also know that if our brains were basically mathematically hard wired every betting company on this earth would have to close down.

  17. Rodger C says:

    2) 3.14159 meters.

Speak Your Mind