Meet the voices: how Luuqad records real Somali speakers

Tap any word in Luuqad and a person says it. Not a synthesized voice, not a recording borrowed from somewhere else, but a Somali speaker who sat in front of a microphone and said hooyo (mother) until it was right. This is how that happens, and why we decided early that it could not happen any other way.

Why a real voice was never optional

Children learn pronunciation by copying. They do not read a word and work out how it should sound; they hear it and say it back, and whatever they hear becomes their accent. If the voice in the app is wrong, the child will be wrong in the same way, confidently, for years.

That puts a lot of weight on the audio. In a language like Somali, where the difference between a short vowel and a long one changes the meaning of a word, a recording that is nearly right is not right. So the rule from the first version of the app has been simple: every word, every phrase, a native speaker, no exceptions. There are 250 entries across the 18 topics, and every one of them has a recording made by a person.

Who does the recording

The voices belong to native speakers of Standard Somali, the Maxaa tiri variety used in schools, on the radio and in the news across the Somali world. We record adults whose everyday speech is that variety, so the child hears the Somali they would meet in a classroom in Hargeisa or a broadcast from Mogadishu, and the Somali that grandparents in the diaspora recognize as correct.

We ask each speaker to do one thing above all: talk the way they would to a child in their own family. Clear, warm, unhurried, but not slowed to the point of strangeness. A word said too slowly is as misleading as one said too fast, because the child learns a rhythm that nobody uses.

Where a word has more than one accepted form, the speaker records the one most widely understood, and the dictionary page on this site notes the alternative. Colors and animals have the most of these, which is why the Colors and Animals pages carry notes on regional usage.

What a synthesized voice gets wrong

It is fair to ask why we do not simply use text-to-speech. It would be cheaper and faster. The answer is that synthetic voices, even good ones, mispronounce Somali in ways that matter.

The first problem is vowel length. Somali writes long vowels with two letters: aabo (father), buug (book), geel (camel), libaax (lion). The length is not decoration. It separates words that are otherwise identical, such as bar (teach) and baar (inspect). Synthetic voices trained mostly on English and other European languages tend to flatten long vowels toward the short ones, so hooyo comes out clipped and saddex (three), with its doubled consonant, loses its shape. We wrote a whole post about this, Long vowels: the one Somali rule that changes everything, because it is the most common error in diaspora Somali.

The second problem is the three consonants English does not have. The letter c is a deep sound made in the throat, as in cun (eat) and cagaar (green). The letter x is a breathy h from the same place, as in madax (head) and lix (six). The letter q comes from the back of the mouth, as in qorrax (sun) and qabow (cold). A synthetic voice that has never been taught Somali maps these onto the nearest English sound, so c is read as an English k or s, x becomes an ordinary h, and q becomes k. A child copying that voice would learn three letters wrong at once. Every one of the 21 consonants is explained, with audio, on the Alphabet page.

The third problem is pitch. Somali uses the rise and fall of the voice to carry meaning that the spelling does not show. The letters i-n-a-n spell both inan (boy) and inan (girl), and only the pitch tells them apart. A person handles this without thinking. A machine, so far, does not.

The checklist every take goes through

A recording is not finished when the speaker stops talking. Each take is listened to against a short checklist before it goes anywhere near the app.

• Long vowels are audibly long and short vowels are short: hooyo, ayeeyo (grandmother), miis (table), suuq (market).
• Doubled consonants are held: saddex, toddoba (seven), siddeed (eight), jaalle (yellow).
• The c, x, q and dh sounds are made the Somali way, not the English way: caano (milk), xanaaqsan (angry), qalin (pen), dhag (ear).
• The pace is even and the last syllable is not swallowed.
• There is no noise behind the voice, and the word starts and ends cleanly so it can be trimmed.

Phrases get one more check. The seventeen greetings in Small Talk and the sixteen questions in Asking are recorded as whole phrases, not stitched together from single words, so Subax wanaagsan (good morning) sounds the way a person says it, with the words flowing together.

Re-recording is normal

Plenty of takes fail the checklist, and that is fine. A vowel is rushed because the speaker knew the word too well. The voice drops at the end of a phrase. A door closes in the next room. The fix is always the same: say it again.

We would rather re-record a word three times than ship a take we are not sure about, because a recording in a children’s app gets played hundreds of times by the same child. Whatever small flaw it carries is repeated that many times too.

From the studio to your child’s ear

The path of a single word looks like this. It starts as a line in a word list, checked for standard spelling. A speaker records it in a quiet room. The take is reviewed against the checklist and re-recorded if needed. It is trimmed so it starts the instant the child taps, and its volume is matched to every other word so nothing is suddenly loud. Then it is placed behind a picture in the Learn tab.

From there it does several jobs. In the Learn tab, the child taps the illustration and hears the word. In the Answer tab, the quiz can play the sound and ask for the picture. In the Play tab, Listen and Tap turns the same recording into a game. And on this site, every entry in the dictionary plays the same audio, so a parent can check a word on a laptop while the child uses the app.

What to do with this at home

Play the word, then say it together. Then ask your child to say it to ayeeyo on the next call and see whether she understands without help. That last test is the one that matters. If a grandmother in Burao or Garowe understands a child in Toronto on the first try, the recording did its job, and so did you.

Written by Luuqad team. The people who make Luuqad. Illustrations from the Luuqad app.