How Well Does Microsoft Speech-to-Text Understand Indian Speech?

When nobody said a time, it writes a clock anyway.

A clock-time opportunity is a clip whose reference contains no time word and no clock — a clip in which nobody said a time. There are 154,555 of them across the eight languages this system covers. It writes a colon-formatted clock on 3,290 of them. The six scored comparators, on the same clips, write one 0 times.

It is not a mis-hearing. A number word is present in the audio — “two”, “a”, “a minute” — and a post-processing stage decides that a number standing near another number must be a time, and formats it as one. The minutes it prints were never spoken and appear in no accepted spelling of the reference.

Ordinary words come back as digits.

Microsoft's model keeps returning digits in sentences where no number exists. These are not misread numbers — they are ordinary words, replaced by numerals. On this test it happens 6,028 times. Across the six comparators combined, it happens 4 times.

"No number exists" is meant literally. A case counts only when a digit is simply not a possibility: the ground truth lists no digit spelling for the word, the word itself is not a number, and no digit is accepted in the slot on either side — so the numeral cannot have spilled over from a real number standing nearby. Every one of the 6,028 is a digit the model introduced on its own.

The mechanism is mishearing toward the nearest number. Urdu بہتر, "better", sounds like bahattar — seventy-two — and comes back as 72. Urdu اسی, "that same", sounds like assī — eighty — and comes back as 80. Malayalam ഒരു, the indefinite article "a", comes back as 1. Speakers naming their own language, छत्तीसगढ़ी, get 36 — the chhattīs hiding inside the name. And this is not one language's quirk: the same substitution runs through Urdu, Bengali, Malayalam and beyond. The words are ordinary. The digits are invented.

Close to one English word in three comes back in a form nobody accepts.

A loanword is a word that belongs natively to English, however often everyday writing renders it in another script. The ground truth is generous with these: it carries optional spellings, so any accepted form of the word counts. Even so, across 187,358 spoken loanwords, Microsoft returns a form that exists nowhere in the lattice 29.03% of the time — close to one word in three. The rest of the field: 10.17–14.46%. These are not technical terms. They are "ma'am", "family", "design".

A tail of words it nearly always gets wrong.

2,310 distinct words are spoken at least ten times, wrong for Microsoft more than 40% of the time, and wrong for the best comparator less than 10% of the time. Together they account for 42,383 word errors. The failure is lexical rather than acoustic — the same word fails repeatedly, from different speakers, on audio the field transcribes.

Some words simply never appear.

110,450 spoken words — 4.79% of the words the reviewers scored — appear nowhere in Microsoft's transcript for that clip: not in place, and not displaced elsewhere. The best comparator loses 1.24%, the worst 2.85%.

Audibility is not the explanation. Narrow the count to words Microsoft alone loses — words that at least five of the six comparators transcribed correctly from the same audio — and 57,140 remain. 52.1% of them are content words of four characters or more. Short function words are 26.9%. Three-character words are 19.1%.

Nor is it one word at a time. 6.2% of its deletion events remove three or more consecutive words, and those runs account for 19.8% of everything it loses.

When a negation is lost, the sentence says the opposite.

A negation is the one word in a sentence that reverses it. Microsoft is wrong on 16.73% of the 45,008 negations spoken, against 4.87–11.44% for the field — but the split matters more than the total.

5.98% of negations are dropped: the word is not written anywhere in the transcript, and the sentence that comes out asserts what the speaker denied. The best comparator drops 1.03% — Microsoft loses a negation 5.8 times as often.

A further 10.76% are replaced by a different word, which damages the sentence in a way a reader cannot detect, because what comes out is fluent.

Signal conditions worsen and yet Microsoft moves less than anyone.

On the 8-language audit corpus, Microsoft posts an overall OI-WER of 23.54% — while the field ranges from 8.38% (Sarvam Audio) to 14.40% (Gemini 3 Flash). Averages hide shape, though. Zoom into Hindi (where per-comparator per-quartile numbers are available in the audit), and Microsoft's *slope* of degradation across audio-quality quartiles is the shallowest of any comparator: a 1.43× noisy-to-clean ratio against a field range of 1.60×–1.87×. Utterance length shows the same pattern — 1.63× for Microsoft, 1.81×–2.99× for the field. Microsoft is the most stable system in the field — but at a floor the leaders never touch. Even on its cleanest Hindi quartile (6.49%), Microsoft trails Sarvam's noisiest (5.24%) — the floor above the leader's ceiling.

Any two systems in the release, district by district.

The district release scores every system on the same audio, district by district, in every language. Pick a language and a system for each map and the two render on one shared colour scale, so the darker map is the weaker system — Microsoft Speech-to-Text against Gemini 3 Pro by default, but any of the twenty-two systems against any other, in any of the fifteen languages. Hover a district to read both error rates and the gap between them.