Catch the miscue mid-word. REALTIME · /p ə n iː m/
A child says "wabbit" instead of "rabbit." A cloud word recogniser only knows once the sentence ends — and it never even notices the substitution.
Pramiteeh's joint word + phoneme stack streams on-device:
every phoneme is scored within ~80 ms of being produced, so the /ɹ/ → /w/ miscue is flagged
while the child is still speaking.
The tutor can interrupt, model the sound, and re-prompt before the child finishes the sentence.
Reading prompt
ReaderAanya, age 6
Sessiongrade-1 fluency
SESSION TIME
0.00 s
Cloud Word ASR
Google / Whisper-style
IDLE
— silent until end-of-utterance —
Tutor signal —
·No analysis yet.
Word + Confidence
Dictation-grade ASR
IDLE
— silent until end-of-utterance —
Tutor signal —
·No analysis yet.
Real-time Word + Phoneme
Pramiteeh on-device · streaming · ~80 ms
IDLE
2.45 s
⚠ live miscue detected
Tutor signal —
·No analysis yet.
Why real-time phonemes change the product
A reading tutor isn't a quiz — it's a coach. A coach corrects during the rep, not after.
Capability
Cloud Word ASR
Word + Confidence
Pramiteeh Streaming Phoneme
Catch a miscue mid-wordbefore the child moves on
no — needs end-of-utterance
no — needs end-of-utterance
flagged within ~80 ms
Detect substitution/w/ for /r/ in rabbit
word still maps
low confidence flag only
explicit /ɹ/ → /w/, live
Detect cluster reductionquickly said as kickly
no
word still recognised
live · /w/ omission
Per-phoneme correctness scoreactionable for IEP / progress tracking
no
word-level only
/0–1/ per phoneme · streaming
Distortion vs. substitution"wrong sound" vs. "weak articulation"