Accuracy & Performance

This page states what has actually been measured, with the raw method behind each number - not marketing claims. Today that is two features: the maths-equation engine (by text and by real voice) and voice typing itself. Everything else is listed below with an honest "not yet measured" note rather than an invented figure.

Azadi Math - equation dictation accuracy

Bangla itself scored 20/20 - perfect, one of 35 languages with a flawless result.

Methodology

A text-door test: 20 real equations (calculus, integrals, quantum mechanics, matrices, chemistry, series and limits) were sent as written spoken-form transcripts, in 70 languages, through the live production endpoint (azadi24.com/api/ai/complete) and checked by the actual shipped OMML converter - not a mock. It measures language understanding and equation accuracy; it does not include the audio/speech-recognition step. Raw per-language and per-equation results are committed in the source repository.

This was a text-door test: it sent written spoken-form transcripts (what someone would say out loud), not real audio - so it measures language understanding and mathematical accuracy, but excludes the speech-recognition step itself (accent, background noise, microphone quality). It ran against the real, live production endpoint and the exact converter shipped in the app - never a lab-only mock.

And spoken end to end - real voice, live in Word

The text bench above deliberately excludes the audio step, and an independent reviewer rightly asked for a test that includes it - with unfamiliar equations. On 2026-08-29 we ran exactly that: 20 equations, SSC to advanced, each dictated by real voice into Microsoft Word on the live production system, one hold per item, every call log-verified. Each equation was spoken TWICE - once following the spoken command words, once phrased casually, the way a teacher speaks in class.

The run also settled claims in both directions. Refuted: spoken pH = -log[H⁺] kept its brackets in three different phrasings, and “pi by two” produced π/2 with no stray 2. Confirmed: a variable spoken immediately after a subscript (the k_B T shape) landed inside the subscript, and the K_c shape dropped its concentration brackets in both phrasings - the run's one consistent weakness class. The full command list and the say-it-this-way tips this produced live on the equations page.

Result by equation

Each of these 20 equations was tested in all 70 languages (max 70 correct per row). The hardest cases involve deeply nested notation - the Schrodinger equation and a chemistry reaction were the two lowest scores, both still above 70%.

Result by language

All 70 tested languages, out of 20 equations each. "Minor" means a small, cosmetically-off but mathematically-readable result; "wrong" means a genuinely incorrect equation.

Voice typing: two modes, two accuracy profiles

Free and Pro voice typing are not the same experience, and stating one accuracy number for both would be misleading. There are two distinct modes:

• Polish mode - hold the shortcut, speak, release: the whole utterance is cleaned up by our AI (punctuation, grammar, mixed Bangla-English handled together) before it is typed. This is Azadi 24's most accurate mode. Pro accounts get it unlimited; a free account gets a daily ration of polish-mode dictations - once they are used up a notification appears, and holding the shortcut falls back to the mode below until the next day (or a Pro upgrade). The current number is on the pricing page.
• Standard streaming mode - the mic-button typing: words appear in real time as you speak, free and unlimited for everyone, forever - it is what makes Bangla voice typing free-forever possible at all. It does not run the AI cleanup pass, so punctuation and mixed-language handling are lighter-touch than polish mode.

Voice typing accuracy - measured

One speaker read 25 mixed Bangla-English test sentences (everyday, business, technical, numbers, a deliberate self-correction). English sentences were read twice - once accent-spoken, once clean - and a few others were retried; one representative take per sentence was scored. Each sentence was said once via polish mode and once via standard streaming mode, on the real shipped app - never a lab mock. Accuracy is WORD-LEVEL: the actual transcribed output was aligned against the known script text word-by-word (a computed edit-distance alignment, not a sentence-by-sentence pass/fail judgement), so one wrong word in a long sentence costs one word, not the whole sentence. Punctuation, capitalization and digit-script/format choices (spelled-out digits vs numerals, comma-grouped amounts) are normalized before comparison and NOT counted as word errors - those are the disclosed, systematic cleanup-depth difference between the two modes (see the section above), not mishearings. One sentence was not dictated in the streaming round (24 of 25 scored there, 466 reference words). Latency is read directly from the app's own logs: polish is release-to-text per call (33 calls); streaming is time from starting the mic to the connection going live (27 sessions) - a different kind of number from polish's, not directly comparable, since streaming then types live with no further per-word wait. This is one speaker, one session - a real spot-check, not a large-scale study.

Test date: 2026-08-28. Polish: 33 calls, latency range 1.2s-3.3s. Streaming: 27 sessions, connection range 2.4s-3.6s.

What actually differed

• Polish mode correctly converted spoken digits to numerals ("একটা নয় সাত আট..." → "১ ৯ ৭ ৮ ০ ৫ ৪ ৩ ২ ১") and formatted large amounts with separators ("৩৪,৭৫০ টাকা"); streaming kept digits closer to the spoken words ("একটা নয় সাত আট শূন্য...") and was inconsistent about digit script elsewhere - a formatting difference, not scored as an error.
• Polish mode's self-correction rule ("say only the final version") did not fire on a deliberate restart in this run - the output kept both the wrong and corrected day instead of the corrected one alone. Every individual word was still heard correctly, which is why this doesn't show up in the word-accuracy score - it's a separate cleanup-logic finding, not a mishearing.
• Streaming mode's word-level errors clustered on English loanwords inside Bangla speech - "endpoint" heard as "in point", "fetch" as "face", "bug" as "বাদ", "launch" as "লান্স" - exactly the kind of mixed-language content polish mode's AI pass is built to clean up.
• Neither mode truncated or summarised the longest test sentence (40 words, about 31 seconds of audio) - both typed the full thing, though streaming still made word-level mishearings within it.

Not yet measured

These features are real and shipped, but do not yet have a published accuracy figure - re-testing them rigorously needs new measurement tooling and real paid API calls, which has not been done yet. We would rather say this plainly than publish an invented number.

• AI-polish output quality on TEXT already typed (grammar, cleanup) - distinct from the voice-typing accuracy measured above
• Translation accuracy
• Latency percentiles for other features (equations, AI writing tools)

Read more about how equation dictation works on the feature documentation page.