Vaaani - The speaking coach that measures your voice, not guesses it

Vaani is an 18-layer acoustic pipeline for IELTS/TOEFL Speaking — Praat formants, peer-reviewed rhythm, syntax-tree grammar → Contrastive Interference Function against 8 Indian-L1 attractors (Bengali + Hindi on Svarah; six on published phonetics). No LLM. No GPU. Same audio = same band. Every L1 callout needs two-piece evidence: acoustic detection AND catalogue match — otherwise unlabelled, never pinned to an L1. Vaani tells you what it measured, how confident, and what it couldn't.

Hi Product Hunt — Neil Shankar here, the maker. I'm an applied linguist, and Vaani started from a frustration I kept seeing in the IELTS / TOEFL prep market: every speaking-prep app gives Indian candidates a single pronunciation number and a vague "work on your accent" hint. Nothing in the pipeline knows their L1. A Bengali speaker whose /θ/ slips to /t/ has a different articulatory fix than a Hindi speaker whose retroflex /ʈ/ bleeds into English /t/ — but the generic scorer treats both as the same mistake. So Vaani measures your voice the way a phonetician would in a lab: Praat for formants and voice quality, two-pass per-speaker pitch tracking, rhythm metrics, Whisper for word-level transcription. Those measurements are compared against an L1-specific acoustic fingerprint, and the report names the exact transfer pattern that produced your band plus the articulatory adjustment that would close the gap. Every band traces back to a measurement you can see in your own report. No LLM in the band-mapping loop — same audio, same band, every time. A few honest constraints I want surfaced on launch day rather than buried: 1. Pronunciation band only. Fluency, Lexical Resource, and Grammatical Range require a human examiner and are explicitly marked "Not scored" on every report. The refusal is the design, not a missing feature. 2. Six L1 profiles — Bengali, Hindi, Tamil, Telugu, Marathi, Gujarati. Bengali and Hindi attractors are empirically calibrated against the Svarah corpus (AI4Bharat, IIT Madras); the other four use published L2-phonetics literature as the prior and are flagged as such on every report. 3. Bengali ground truth is cross-referenced against Asoke Kumar Datta's Acoustics of Bangla Speech Sounds (ISI Kolkata / Springer 2017) — the canonical acoustic-phonetic record for Standard Colloquial Bengali. 4. Not an official score. No affiliation with Cambridge, IDP, British Council, or ETS. A diagnostic instrument, not a substitute for an examiner. What evolved over the build: I started with a full four-criterion IELTS rubric and stripped it back to acoustic-core Pronunciation-only after a calibration round showed me that honest reporting beats feature parity. The engine also moved from a GPU-leaning phoneme-alignment stack to a CPU-only Praat-centric one — slower per submission, but more reliable on real Indian-accented audio and far easier to deploy. I'd love feedback from anyone prepping for IELTS / TOEFL Speaking, anyone running a coaching centre in tier-2/3 India, or phonetics-literate folks willing to pressure-test the methodology. Critical questions especially welcome. — Neil Shankar Ray

Vaaani - The speaking coach that measures your voice, not guesses it

Replies