English to IPA, and then hear it said

Both of the ones dictionaries use. The British column follows Received Pronunciation, the spelling in Cambridge and Oxford: bath is /bɑːθ/, car is /kɑː/. The American column follows General American: bath is /bæθ/, car is /kɑːr/. Stress marks sit at the start of the syllable, the way every dictionary prints them, not next to the vowel.

Usually. Those words are spelled one way and said two ways, and the only thing that separates them is the rest of the sentence. Schwa reads the whole sentence rather than looking words up one at a time. On 24 test sentences built in matched pairs it picked the right reading 71% of the time. It is not perfect, and the page lists the pairs it still gets wrong instead of claiming it does not miss any.

Out of the 20,000 most common English words, 8,639 come out differently in the two spellings, which is 43.2%. Most of that is one thing: the letter r after a vowel, which is said in America and Canada and dropped in most of England, Australia and New Zealand. The rest is a smaller set of real differences, like the vowel in bath and the first syllable of schedule.

We measure it instead of trusting the label. Every speaker in Common Voice states their own accent, but a voice built from a recording can lose it, or pick up one that was never there. So each voice reads a sentence full of words with r after a vowel and a control sentence with none, and both are measured for the resonance that drops when an English r is curled. 13 voices came out r-coloured and 15 did not. 5 of the 31 read against their own accent, mostly by acquiring an American r the speaker never had, and those are not on the site.

25 English voices, 11 women and 14 men, from 25 different people across 9 places where English is a first language. They are built from Mozilla Common Voice recordings, which are CC0 1.0, public domain, with no conditions attached. They are not one voice at different pitches.

Not for lack of recordings. Indian and South Asian English is the third largest group in the whole English corpus, ahead of six of the accents that are on the site. The site draws its line at communities where English is the first language of the community itself, because that is the only line here that is not arbitrary, and because the phonetic spelling the page prints is British or American. Drawing it elsewhere would be a different site, and a good one.

Yes. Up to 500 words (3,900 characters) at a time, with 5 seconds between two goes. No account, no card, no watermark on the MP3. You can use it in work you are paid for.

Yes, 64 more, and each one is read by a person who speaks it, not by an English voice with a foreign accent. 30 of them have their own page with a sample you can hear before typing anything.

It is not stored. The server keeps only a checksum, so that if the same text is asked for twice with the same voice it can hand back the finished file instead of using the graphics card again. Audio files are deleted after 24 hours.