The tune under the words

It is said that when Mamman Shata was at a loss for words, his kalangu players could help him. A drum has no letters, only the pitch and rhythm of speech, and in our languages that carries more of the meaning than writing shows.

The Hausa singer Mamman Shata made up his songs as he sang them, in front of the people he was singing for. He lived from 1923 to 1999, and he performed with at least two players of the kalangu, the hourglass talking drum of Hausa music. It is said that when Shata was at a loss for words, his drummers could help him.

The accounts do not say how, and we will not invent it. We do know what a talking drum carries. It copies two things from speech, how high each syllable sits and how long it lasts, and it makes no vowels or consonants of its own. Hausa musicians hold that their instruments talk, with phrases of the language tied to particular sounds, and Hausa has a word of its own, take, for the praise names played on a drum. In Songhay, the language family of Zarma, the talking drum is the doodo. So perhaps the help was the shape of the next line, its rises and falls, arriving on the drum before the words did.

It works because in Hausa the tune is part of the word. The language has a high tone, a low tone and a falling tone, and no rising tone. Here are three words that look different on paper: karan, the stalk; karen, the dog; karon, the collision. Said inside a phrase, they come out with the same consonants and vowels, and only the pitch tells them apart. The linguist Russell Schuh called them a minimal triplet for tone. Anyone raised on Hausa hears which one was meant without stopping to think.

The page and the voice keep those three words apart in different places. On paper the vowels do it, a, e and o. In a phrase those vowels run together and the tune takes over. Sometimes it goes the other way. Hausa's everyday spelling in Latin letters marks neither tone nor the length of a vowel, and references on the language give the example of daga, one written word for two spoken ones. Said low and short, it means "from". Said high, with long vowels, it means "battle". A reader puts the tune back from memory, the way someone who knows a song can read its words and hear it.

Zarma has a tune of its own. It is usually described with four tones, high, low, falling and rising, the last one a tone Hausa does not have, and they too mostly stay off the page.

The tune also changes from place to place. The linguist Paul Newman described what Hausa speakers in Niger do when two tones meet and the voice has to simplify them. In Ader they settle it one way and in Dogondoutchi another, while in Maradi people do both, depending on how long the word is. A sentence can travel from one place to the next on a slightly different melody and be good Hausa in each.

None of this is a puzzle for the people who speak these languages. It is ordinary listening, the kind a child picks up long before school. It is also what a machine misses when it learns a language from text alone, since written Hausa and Zarma usually carry no tones.

For a machine to hear the tune as well as the letters, pitch has to count as part of the word. It would tell the dog from the stalk by the voice alone, as a Hausa listener does, and it would take the Hausa of Maradi and the Hausa of Dogondoutchi as two right ways through the same tones. Speaking asks for the same knowledge in reverse. A machine that talks has to put the tune back into words that were written without one, and if it guesses wrong, daga can come out as "battle" in a sentence that meant "from".

This is part of our work at Namu. We build speech AI so that people in the Sahel can use technology by speaking, in their own language, and the tune is a large part of what that AI has to learn to hear and to give back.

When Schuh wrote the three words down, he put small marks over the vowels for the low and falling tones. Rub out the letters and keep only the marks, and what is left is a short row of highs and lows. A kalangu player could play it.