Bodhi understands India best — in all nine languages.

The standard test · word error rate

Everyone scores speech-to-text the same way: count the wrong words.

There is one number the whole industry lives by — Word Error Rate, or WER. It's simple: take what the system typed, line it up against what the person actually said, and count how many words are wrong, missing or invented. Divide by the number of words spoken. Lower is better — a WER of 12% means roughly one word in eight came out different from what was said.

It's the number on every vendor's slide, ours included. So it's the honest place to start: measure every system the same way, on the same audio, and see who gets the most words right. We did exactly that across nine languages. But before the scores, the catch that this number hides — because it's the whole reason a good-looking WER can still let you down.

The catch · what the number hides

A low error rate can still fail you on a real call.

WER counts every wrong word the same. But not all wrong words matter the same. Writing “colour” for “color”, or “25” for “twenty-five”, counts as a mistake — even though nobody is confused. Now compare that to mishearing a customer's name, flipping a “yes” into a “no”, or hearing a different amount. Those count as one mistake each too — yet only these end a conversation. The number treats a harmless spelling quirk and a wrong name as exactly equal.

So a system can post a respectable error rate and still be the one that mangles the name on every call. To judge what a system is really worth, you have to look at the kind of mistake it makes, not just how many. Here is one clip that makes the point — a plain farming question in Bengali:

The results · Bodhi vs the field

Even by the raw number, Bodhi leads all nine languages — with 11–22% fewer errors than the next-best system, every time.

Knowing WER's blind spot, we still measured it — the same clean-up applied identically to every system, on the same clips — because it is the common yardstick everyone recognises. Across about 67,000 clips from six public datasets, Bodhi has the lowest error rate in every one of the nine languages.

Here is Bodhi against its closest competitor in each language — whichever of the four rivals scores best there, named on its bar. Bodhi's bar is shorter every time:

And the margin is not thin. Language by language, Bodhi makes 11% to 22% fewer mistakes than the closest competitor — roughly one error in six avoided, on average. To be clear about what this counts: it is the raw tally, where a harmless spelling variant and a wrong name each cost one point. Whether the mistakes a system does make are the kind that change the meaning is a separate question — the second test, further down, answers it. And that is against the best rival each time, which changes from language to language; measured against the rest of the field the gap widens sharply — in six of the nine languages, the weakest of the four competitors makes more than twice as many mistakes as Bodhi. All five systems, side by side:

Every language here — Hindi included — is measured the same way, on the same public datasets, with the same scoring. Bodhi has the lowest error rate in all nine.

The evidence · hear it yourself

The mistakes behind the numbers — one clip per language.

Numbers are easy to argue with; recordings aren't. Below is one real clip per language where Bodhi got the sentence right and a rival changed the meaning — a wrong word, a wrong name, a dropped word. Press play, read what each system heard, and follow the plain-English note on what went wrong. (Hindi's own worked examples are in our dedicated Hindi report.) Pick a language tab:

Test 2 · did it keep the meaning?

Then we checked the harder thing — the meaning. No system keeps it better than Bodhi.

Word error rate can't tell a harmless spelling quirk from a wrong name. So we ran the test that actually matters for a voice agent: a blind AI judge — OpenAI's GPT-5.4-mini — read the human reference and every system's transcript, with brand names hidden and the order shuffled on every clip, and rated how faithfully each one preserved what was actually said. About 4,700 clips across the eight non-Hindi languages — with Hindi judged the same way separately (just below).

Clip for clip, no system keeps the meaning better than Bodhi. Judged head-to-head on the same audio, Bodhi comes out ahead of every competitor:

Bodhi vs each rival on the same clips, rated by the same blind judge · numbers are clip counts.

And on the headline measure — the share of clips where the meaning came through intact — Bodhi sits at the top of the field:

Hindi tells the same story. In our dedicated Hindi study, judged the same blind way, Bodhi kept the full meaning on 72% of clips; the next-best of these four systems, Sarvam, managed 59% — and Bodhi wins every Hindi head-to-head too.^*

^* Hindi figures from our dedicated Hindi report, judged the same blind way.

Why the gap · data and design

Why does Bodhi come out on top?

There is no single trick. The gap you see above is the sum of building for Indian speech from day one, rather than adapting to it later:

Trained on the speech India actually speaks. Bodhi learns from real Indian conversations — telephony-grade audio with traffic and crosstalk behind it, regional accents and dialects, and the code-mixed English (phone, payment, slip) that formal corpora underrepresent. Most global systems meet this speech for the first time in production; Bodhi grew up on it.

Dialect coverage, not just the textbook language. No Indian language is spoken one way — the Hindi of Patna is not the Hindi of Jaipur, and Tamil, Bengali and Kannada shift accent, vocabulary and phrasing from district to district. A system trained on the news-reader “standard” starts failing the moment a caller sounds like where they're from — which is most callers. Bodhi is trained and evaluated across these regional varieties, and the benchmark you just read reflects that: it includes dialect-rich corpora such as RESPIN and Kathbath, recorded from speakers across districts and states — so the accuracy holds in the field, not just in the metros.

Tuned for the errors that end conversations. A voice agent acts on names, numbers and negations, so that is what we optimise hardest — it's why, in the clips above, rivals lose a person's name or an amount where Bodhi holds on.

Indic scripts as first-class citizens. Nine languages, three script families, and dozens of equally-correct ways to spell the same word. Bodhi is built around that reality — spelling variants, joined-vs-split compounds and both scripts of code-mixed speech are modelled, not fought.

Measured the way you just read. Benchmarks like this one run internally, per language, on every model iteration — so a win in one language is never bought with a regression in another.

And we put our work back into the Indian voice-AI space. This entire benchmark — the audio, the human references, every system's transcripts and the blind judge's ratings, across ~67,000 clips in nine languages — is openly published on Hugging Face, alongside our dedicated Hindi benchmark, for anyone building or evaluating Indian-language voice AI to use. And we build for the languages the industry skips: Bodhi serves Odia and Malayalam callers that some global providers don't support at all — because a voice AI for India has to work for all of India.