Learn Sinhala by Speaking: Why Most Apps Get Sinhala Wrong

Sinhala has four prenasalized voiced stops — a short nasal fused directly onto a following stop consonant, shorter than a full nasal sound and shorter than a nasal-plus-stop cluster too. It's its own distinct sound, not a fast version of either. This particular contrast — prenasalized consonants standing opposite genuine nasal-plus-stop clusters as two separately meaningful sounds — has been reported in only three languages worldwide: Sinhala, and the unrelated Fula (spoken across West Africa) and Selayarese (spoken in Indonesia). Most speech models have never encountered it in training data, because almost nothing they were trained on has it.

Sinhala has a genuine four-way contrast built around a single syllable: කද (kadə, "shoulder pole") uses a plain voiced stop with no nasal at all; කන (kanə, "ear") uses a plain nasal with no stop; කඳ (kaⁿdə, "trunk") uses a short, fused prenasalized stop; and කන්ද (kandə, "hill") uses a full nasal-plus-stop cluster. Four real, unrelated words, distinguished only by how a nasal element does or doesn't attach to a following stop, and for how long. English has nothing that maps onto this at all — English speakers have never had to listen for the difference between "a nasal fused onto a stop" and "a nasal followed by a stop," because English doesn't treat that distinction as meaningful.

The script actually marks this precisely: Sinhala has dedicated single letters for the prenasalized consonants (ඹ for the prenasalized b, ඳ for the prenasalized d, and others), separate from both the plain stop letter and the letter combination used to spell a genuine nasal-plus-stop cluster like the one in කන්ද. The spelling distinction exists. What doesn't survive into a transcript is the actual duration and articulation that makes the sound recognizable as prenasalized in the first place — which is exactly the part a learner's mouth has to get right, and exactly the part a text-matching app never checks.

Sinhala's script still carries a full set of separate letters for aspirated consonants, inherited from Sanskrit and Pali spelling conventions — the kind of letter that, in Hindi or Punjabi, marks a real, meaning-changing puff of air after the consonant. In modern spoken Sinhala, that distinction is gone. A word spelled with an aspirated letter and its unaspirated counterpart are pronounced identically; the extra letters survive only to spell Sanskrit- and Pali-derived vocabulary correctly, not because anyone still says them differently. කනවා (kanavā, "to eat") corresponds to a Sanskrit root that had aspiration; the modern Sinhala pronunciation doesn't carry it.

That's close to the reverse of what an app built around general South Asian language assumptions would expect. Aspiration is a load-bearing, meaning-changing feature in Hindi, Punjabi, and Bengali, so it's a natural thing for a model or a learner to assume matters everywhere in the region. In Sinhala it's a spelling artifact, not a pronunciation target — while the sound that actually is meaning-changing, prenasalization, isn't the kind of distinction those neighboring languages train an ear (or a model) to listen for at all.

Prenasalization is a timing and articulation property — how long the nasal portion lasts, and how tightly it's fused to the following consonant — and that kind of detail doesn't survive being reduced to text. Even with the script marking the distinction, a system that works from a transcribed or auto-recognized rendering of your speech can mistake one for the other under real, imperfect pronunciation, because the two sounds sit close together in duration. The distinction that actually changes the word from "trunk" to "hill" is exactly the kind of acoustic detail a transcript-first pipeline is most likely to blur.

Layer the aspiration mismatch on top and the problem compounds: a system tuned on the assumption that South Asian languages care about aspiration might spend feedback cycles correcting a distinction Sinhala speakers don't make, while having no mechanism at all for the timing-based contrast that actually carries meaning.

Audio-level feedback on timing, not just which letter was produced. Prenasalization versus a full nasal-plus-stop cluster is a duration difference in the actual sound, not something a transcript can distinguish.

A model that doesn't import assumptions from neighboring languages. Sinhala dropped a distinction (aspiration) that's very much alive in Hindi and Punjabi, and kept one (prenasalization) that almost no other language in the world has. Correct feedback depends on treating Sinhala's actual sound system, not a generic South Asian template.

Real spoken examples, since a contrast this rare — reported in only two other languages on Earth — isn't something most learners have any prior intuition for at all.

Native audio processing, not transcript matching. Yapr's speech-to-speech pipeline listens to the actual duration and articulation of your speech, so it has a chance at telling a genuine prenasalized stop apart from a full nasal-plus-stop cluster instead of treating both as the same transcribed letters.

Feedback tuned to Sinhala specifically. Aspiration doesn't carry meaning in spoken Sinhala the way it does in Hindi or Punjabi — Yapr doesn't waste your practice time correcting a distinction that isn't real here.

Real conversation, not isolated drills. A contrast this unfamiliar is easiest to internalize by hearing it used naturally, over and over, in context.

Whisper mode. Practice an unfamiliar sound like the prenasalized stop privately before trying it out loud with someone who'll notice.

$12.99/month, 47 languages. One subscription covers Sinhala and anything else you're learning, against the cost and scheduling friction of finding a Sinhala-speaking tutor.

Yapr processes what you actually said in Sinhala, including the timing that separates a prenasalized consonant from a full nasal-plus-stop cluster, instead of just checking whether your transcript matches the text on the screen. 47 languages at yapr.ca.