Can AI really hear my th sound, or does it just guess from context?
Dictation apps guess TH from context and autocorrect it. A phoneme recognizer scores the sound itself. Here's the difference, and which one you're using.
Yes, but only if the app is built to listen for the sound itself rather than to guess the word you probably meant. That distinction is the whole answer, and it is worth understanding before you trust any pronunciation score.
30-second version: A transcription model (the kind behind dictation and most "pronunciation" apps) predicts the most likely word, and its language model quietly repairs a mispronounced TH back into the correct spelling. A phoneme recognizer predicts the most likely sound, frame by frame, with no vocabulary to fall back on. Only the second kind can actually tell you that you said "sink" instead of "think."
Why context-guessing apps miss it
Automatic speech recognition was built to solve dictation: turn speech into readable text as fast and accurately as possible. To do that well, these models lean hard on a language model — statistics about which words tend to follow which other words. That's a feature for dictation and a liability for pronunciation checking, because a language model's whole job is to smooth over exactly the kind of error you're trying to catch.
Say "I sink about this a lot" clearly enough, with a crisp /s/ where /θ/ should be, and a transcription-based system has two candidate words in front of it: "sink" and "think." "I think about this a lot" is overwhelmingly more common in its training data than "I sink about this a lot," so it writes down "think" — and if the app then congratulates you for saying "think" correctly, it has scored your intention, not your mouth. This is the same failure mode covered in the autocorrect problem: a dictation engine is optimized to forgive you.
How sayit catches it instead
sayit's phoneme layer runs a wav2vec2-family acoustic model that outputs a probability for every English sound in each fraction-of-a-second slice of your audio — not a transcript, a running estimate of which sound it's hearing. There's no dictionary sitting on top deciding that "think" is more likely than "sink," because the model isn't choosing between words at all. It's reporting what it heard, sound by sound, then that sequence is aligned against the canonical IPA for your target word.
For a word like think, the target phones are /θ ɪ ŋ k/. If your recording produced /s ɪ ŋ k/, the alignment shows a substitution at the first slot — target /θ/, produced /s/ — and the word gets flagged with a confidence-based verdict rather than a silent pass. Open the word and you see the two phone sequences side by side, plus a "sounded like" hint if what you produced happens to spell a real word ("think" → "sink" is exactly that case).
What the fix looks like
Once you can see the substitution named — /θ/ produced as /s/ — the fix is mechanical rather than mysterious. /θ/ needs the tongue tip lightly between the teeth with continuous airflow; /s/ is made further back, tongue near the ridge, no tongue-teeth contact at all. Practise the contrast directly rather than the word in isolation, because in isolation your ear may not be able to tell the two sounds apart yet — that's a large part of why the substitution happened in the first place.
Drill the minimal pair think vs sink, say the pair fast a few times, then record a full sentence and check the same word again. The full /phonemes/th-voiceless page has the articulation steps in more depth than fits here.
How do I know it's actually improving?
Track the same word or sentence across a few days rather than trusting a single take. A confident /θ/ shows up as a good verdict with high acoustic confidence; a shaky one that's on its way to fixed often shows up as unclear — not wrong, but not yet reliable — before it settles into good. If you keep hearing /s/ named specifically in the breakdown, that's your actual curriculum; if the substitution stops appearing across several different words, the fix has generalized past the drill sentence, which is the real test.
Try the two-second version
Say this sentence and see whether the app named the sound or just wrote down the word it guessed you meant:
Record: "I think the sink needs three thin pipes." It packs /θ/ and its most common substitute into one sentence, five times. Open sayit free — no install, no card — and look at whether "think" and "sink" get told apart.
Hear exactly which sounds to fix.
Say one sentence and get sound-by-sound feedback in seconds. No install, no card.