Naming the picture aloud with “Say the word”
In the “Say the word” activity, the child names the picture into the microphone and Lexikit recognises the word. The sound does not leave the device.
“Say the word” shows a picture and waits for the child to name it out loud. The device listens and recognises the word among those of the resource. When the word is recognised, the next picture appears, and the voice repeats the word with its article: “a rabbit”.
It is the first activity in Lexikit where the child produces the word. In bingo or “Where is…?”, the child recognises it. Naming a picture asks for more: finding the word without hearing it. This step brings the word into the vocabulary the child uses on their own.
Why everything stays on the device
Browsers can already recognise speech, but to do it Chrome sends the sound to Google's servers. I did not want to send children's voices to an outside service. So Lexikit uses Vosk, a free recognition engine that runs entirely in the browser. The language model, a few dozen megabytes, downloads once and stays in the browser. No sound leaves the device and nothing is recorded.
I also tried Whisper, a more recent engine. On a single word, it tends to make up a whole sentence, and it is too heavy for a school tablet.
A closed list of words
Vosk does not search among every word of French. It chooses among the words of the resource, plus a few small words that often come around them: “c'est”, “un”, “une” (it's, a). Without these small words, “c'est un lapin” (it's a rabbit) became “chat lapin” (cat rabbit). With them, a word said slightly wrong is still recognised.
In French, the “Strictness” setting adds neighbouring words that sound almost the same. On Lenient, only the words of the resource count. On Medium, “tapin” or “papin” no longer pass for “lapin” (rabbit). On Strict, neither does “labin”. “Lapain” is accepted at every level, and a clearly spoken “sapin” (fir tree) is refused even on Lenient.
Never “no”
When the word is not recognised, Lexikit says “I didn't hear you well” and lets the child try again. After a second missed try, the child taps the right picture among four. A silence does not count as a mistake.
What remains to be checked
The Vosk models learnt mostly from adult voices. My tests, stammering or mangling the words, work, but the real test is a classroom, with children's voices and noise. The “Strictness” setting exists only in French for now, and the iPad is not tested yet. If you try it with a class, tell me how it went through “Your feedback”.
