The languages it does badly, and who that is

The unevenness, stated plainly

English has thousands of hours of clean, transcribed, permissively available speech. So do a handful of others — Mandarin, Spanish, French, German, Japanese. Below that, availability falls off a cliff.

There are roughly seven thousand living languages. Fewer than a hundred have speech systems anyone would call good. Several hundred million people speak languages with no usable speech recognition at all.

Within a well-served language, the same unevenness repeats across accents. Recognition error rates for a language's prestige accent and for a regional or second-language accent of the same language routinely differ by a factor of two or three. Published audits have found large disparities in commercial systems between speakers of different varieties of English, in the same recording conditions, saying the same things.

This is not a design choice. It is the training distribution, exactly as the caption corpus was for images. But the consequence is not aesthetic here — it is a person being unable to use a voice interface, an automatic transcript of their evidence being wrong, or a subtitle system garbling their name.

What breaks specifically

Code-switching. Speakers of many languages mix in English or another language mid-sentence as a matter of ordinary speech. Most systems are built around one language per utterance and handle a switch by mangling both sides of it. Hinglish, Taglish, Spanglish and dozens of comparable everyday registers are, from the system's point of view, an error condition.

Names. Personal and place names outside the training distribution are transcribed phonetically and wrongly, which is a small daily indignity and occasionally a serious problem in medical or legal transcription.

Tone and length distinctions that English does not make are frequently flattened by both recognition and synthesis, changing meaning rather than merely accent.

Scripts with ambiguous vowel marking — Arabic, Hebrew, and abjads generally — require the system to infer pronunciation that the text does not carry. Diacritics resolve it and are usually absent in ordinary writing.

What is being done, and what you can do

The serious work here is community-driven and worth knowing about because you can join it.

Open, community-collected speech corpora exist that anyone can contribute to by reading sentences aloud, and anyone can use to train. Several languages have crossed the threshold from unusable to usable purely on this basis, contributed by their own speakers.

The practical routes if your language is poorly served:

• Fine-tune an open recognition model on a few hours of your language or accent. This is genuinely achievable — free notebook time is enough — and the improvement on a specific accent from even a small, well-matched set is substantial.
• Use a language-specific model rather than a multilingual one where it exists. The multilingual giants are convenient and frequently worse than a small model trained on one language.
• Contribute recordings to an open corpus. This is slow, collective and the only thing that fixes it durably.
• Check the numbers before you promise anything. Published error rates are usually reported on clean read speech in the best-served accent. Test on your own audio before building a product on it.

Why to be sceptical of "supports 100 languages"

That claim is now standard and it means very little. Support ranges from "trained on thousands of hours" to "trained on a few hundred, mostly religious texts read aloud". The word covers both.

The question that separates them: what is the measured error rate on this language, on spontaneous conversational speech, in ordinary recording conditions? Vendors that have the number will give it. Vendors that quote a single headline figure across all languages are quoting the best one.

The unresolved part, and it is worth naming as unresolved: nobody has a route to good coverage of the long tail that does not depend on the speakers of those languages doing the data collection themselves, largely unpaid. That is a real problem of fairness with no agreed answer, and the current arrangement — that the world's least-served speakers must volunteer to build their own tools — is defended by some as community ownership and criticised by others as free labour. Both readings are available and both are honest.