Why speech recognition doesn't catch pronunciation mistakes

Say "I sink so" into almost any AI speaking app. Watch what happens: the transcript reads "I think so", the app congratulates you on a clear sentence, and you move on. The mispronunciation you most needed to hear about was silently repaired by the machine — and it will still be there in your next job interview, where the listener is not a forgiving algorithm.

This is not a bug in one app. It is a structural property of standard speech recognition, and it quietly undermines an entire category of pronunciation feedback. Understanding it takes five minutes and will change how you evaluate every speaking app you try.

Speech recognition is built to guess what you meant

Modern speech-to-text (ASR — automatic speech recognition, the technology behind dictation, subtitles, and Whisper-style models) is optimized for one goal: produce the most likely text given the audio. Not the most faithful rendering of the sounds you made — the most plausible words.

To do that, an ASR system effectively combines two kinds of knowledge:

• Acoustic evidence — what the audio actually sounds like.
• Language knowledge — which word sequences are probable in English.

That second part is the language model, and it is doing more work than most people realize. When the acoustics are ambiguous — an accented vowel, a substituted consonant, a swallowed ending — the language model breaks the tie in favor of real, likely words. "I sink so" is a wildly improbable English sentence; "I think so" is one of the most common. So the system writes "think". Confidently. Every time.

For transcription, this is exactly right. It is why dictation works in a noisy café and why subtitles survive mumbling actors. ASR engineers spent decades making recognizers robust to pronunciation variation — which is a precise way of saying: trained to ignore the very thing a pronunciation learner needs measured.

The error laundering pipeline

Follow a single mistake through a transcript-based speaking app:

1. 1.You say /aɪ sɪŋk soʊ/ — "I sink so", with /s/ where /θ/ belongs.
2. 2.The recognizer's acoustic front-end hears something between /s/ and /θ/.
3. 3.The language model weighs in: "think" is the only plausible word here.
4. 4.The transcript says "I think so". The app compares transcript to target. Match.
5. 5.Green checkmark. "Great pronunciation!"

The error went in; a compliment came out. We call this error laundering: the mistake is washed clean before any feedback logic ever sees it. The app is not lying to you — it genuinely cannot see the error, because its only view of your speech is a transcript from which the error has already been removed.

And the better the recognizer, the worse the problem. Each new generation of ASR is more robust to accents and slips — which means it repairs more of your mistakes, and the feedback built on top of it goes blinder.

Why conversation apps inherit this

AI conversation apps — the Speak and Talkpal category, and every "talk to an AI tutor" feature — are built on exactly this pipeline: your speech → ASR transcript → a large language model generates a reply. It is a genuinely great architecture for what those apps do best: fluency, confidence, vocabulary, keeping you talking. We recommend conversation practice regularly in our comparison of pronunciation apps.

But notice what the AI tutor actually receives: text. By the time your words reach the model that "understands" you, your pronunciation is gone — laundered into clean text by the recognizer. The tutor could not comment on your /θ/ if it wanted to; it never heard it. When such apps offer pronunciation scores at all, they are typically derived from transcript matching or recognizer confidence, both of which sit downstream of the auto-correction.

This is why learners plateau in a specific, frustrating way: months of daily AI conversation, real gains in fluency — and the same five sound errors as day one, because no part of the loop was capable of noticing them. If people still ask you to repeat words, this is likely why.

The two-minute test

You do not have to take our word for any of this. Test any app you are using right now:

1. 1.Pick a sentence with a sound you know is hard for you — or fake one: "I think the theory is right."
2. 2.Deliberately mispronounce it: "I sink the seory is right." Commit to the error.
3. 3.See what the app says.

There are only two outcomes. Either the app shows you exactly where /θ/ became /s/ — in which case it is genuinely listening to your sounds — or it congratulates you on a sentence you deliberately said wrong. If it cannot catch an error you made on purpose, it is not catching the ones you make by accident.

Run the same test with rice/lice if /r/–/l/ is your challenge, or vest/west for /v/–/w/, or ship/sheep for /ɪ/ vs /iː/. The minimal-pairs guide has plenty of test material.

The fix: take the language model out

The root cause is the language model's veto over the acoustics — so the fix is architectural: remove it.

sayit's recognizer is a neural language-model-free phoneme recognizer. It is never asked "what word is this?" — only "what sound is this, instant by instant?". There is no dictionary pulling /sɪŋk/ toward "think", because the system does not operate on words at all. It outputs the phoneme sequence you actually produced, which then gets aligned against the target sequence and scored sound by sound. The full pipeline — posteriors, alignment, per-phoneme goodness scores, prosody — is walked through in how AI pronunciation scoring works.

The result is feedback with a different character:

• Say "I sink so" and sayit shows sink's first consonant red on the heatmap, with target /θ/ beside the /s/ you actually made, and a concrete tip: tongue tip between the teeth, keep the air flowing.
• Say something in your own accent that is regionally different but perfectly clear, and it passes — hearing raw sounds also means the engine can be deliberately tolerant of accent variation while enforcing contrasts that change meaning.

One honest caveat: a phoneme recognizer is the harder path. Word-level ASR gets to lean on a dictionary to smooth over uncertainty; a phoneme system has to stand behind every sound it reports, which demands a stronger acoustic model and careful score calibration. That is the engineering trade we chose, because for assessment — unlike transcription — being corrected by the machine is the failure mode.

Use both tools for what they are good at

None of this means abandoning conversation practice. The sensible split:

• Conversation apps for fluency, hesitation, vocabulary and confidence — the skill of keeping speech flowing.
• A phoneme-level tool for the sounds themselves — diagnosis and targeted drilling, whether that is structured exercises, the hardest English sounds, or checking a single word in the word lookup hub.

sayit covers the second job end to end: per-phoneme heatmaps, IPA target-vs-actual with articulation tips, stress and intonation scoring, a Magic Wand coach that finds your systematic weaknesses and builds a drill plan, plus reading, shadowing, dictation and freeform modes. The features page has the full picture.

Try the test right now

The two-minute test works on us too — we built the engine to pass it. Open sayit free, say a sentence with one deliberate mistake, and see whether it catches the exact sound. No install, no card; it runs in the browser. If it ever congratulates you on an error you made on purpose, tell us — that is a bug report we want.