Why Turkish dictation gets English words wrong
English terms inside a Turkish sentence usually come back from dictation unrecognisable. The mechanisms below stack up to produce that, and not one of them is a malfunction: each is what Turkish speech recognition is designed to do.
Turkish spelling is phonemic
In the Turkish alphabet a letter largely stands for one sound. A recogniser trained for Turkish has a Turkish output alphabet too: whatever it hears, it has to write with Turkish letters.
English puts far more distance between spelling and pronunciation. The recogniser cannot close that distance, because closing it would first require noticing that the word is English, and in a Turkish session it does not have that information.
“diploy”
deploy
“keş”
cache
“fiçır”
feature
“bild”
build
English vowels round to the eight Turkish ones
Turkish has eight vowels: a, e, ı, i, o, ö, u, ü. English has roughly fifteen vowels and diphthongs depending on how you count them. Every vowel with no Turkish counterpart lands on the nearest one that exists.
The reduced vowel in the unstressed syllable of production does not exist in Turkish; a and ı arrive instead and the word becomes prodakşın. The long vowel in merge drops to an e and merç comes out. This is not sloppy approximation, it is the closest answer the Turkish sound inventory can give.
“prodakşın”
production
“merç”
merge
“ribeyz”
rebase
“mayın”
main
Consonant clusters and final devoicing reshape the word
Turkish syllable structure is narrower than English. Words do not begin with a consonant cluster, so a vowel moves in and stack becomes istek. Voiced consonants harden at the end of a word, so bug becomes bağ.
In both cases the form that comes out is a real Turkish word. That is precisely why, for the reason in the next section, it becomes the candidate the recogniser prefers.
“istek”
stack
“bağ”
bug
“branş”
branch
“kreş”
crash
The language model picks the word that is in its vocabulary
A speech recogniser is not only an acoustic model. A language model sits above it, knowing which word sequences are likely, and that model was trained on Turkish text. Given two close candidates, the one in its vocabulary that occurs often wins.
This is where the funniest entries come from. The sound of staging is close to the Turkish word stajyer, which means intern and is common; the recogniser takes it, and the sentence turns into stajyerde çalışan build. Label is read as a given name and written with a capital letter as Leyla. The recogniser is not making a mistake here. It is returning the likeliest answer within its own vocabulary.
“stajyer”
staging
“doktor”
Docker
“tokat”
token
“çıra”
Jira
“Leyla”
label
“kola”
call
A wrong stem pulls in a wrong suffix
Turkish is agglutinative, and its suffixes attach to pronunciation rather than spelling: the last vowel of the word decides which suffix arrives, which is vowel harmony. Once the stem is written wrongly, the suffix is chosen against that wrong stem and the error grows to the end of the word.
Production is pronounced prodakşın in Turkish, so it takes a back-vowel suffix and the correct spelling is production'da. Cache is pronounced keş, takes a front-vowel suffix, and gives cache'i. Getting this right means holding the English spelling and the Turkish pronunciation at the same time: one decides the letters, the other decides the suffix.
“prodakşında”
production'da
“keşi”
cache'i
“branşını”
branch'ini
“kumiteleri”
commit'leri
Code-switched speech is largely absent from the training data
Speech recognition models are usually trained on monolingual data, or on data split by language. Switching language mid-sentence, which is code-switching, is a recognised and separately studied problem, yet most production systems still pick a single language for an utterance.
The moment Turkish is picked, the fate of the English words in the sentence is settled. Every mechanism above runs downstream of that one decision.
How TrenVoice approaches this
TrenVoice runs a multilingual speech model that treats a code-switched sentence as ordinary input rather than an exception, together with a glossary of terms you approved yourself. The glossary touches only the terms you added and the Turkish suffixes attached to them, and leaves the rest of the sentence as spoken. Dictation runs entirely on your device.