Speech synthesis picks a voice by locale, not by the language of the text

Fact

A SpeechSynthesisUtterance with no lang and no voice inherits the language of the document — and failing that, the host's default voice. Nothing inspects the string being spoken. So English text handed to a browser on a non-English system is read aloud by a voice for that system's locale, which pronounces it with that language's phonetics and produces something the listener cannot parse as English at all.

The obvious fix is half a fix, because speechSynthesis.getVoices() returns an empty array on first call in most browsers. The list is populated asynchronously and announced by a voiceschanged event, so voice selection written inline at startup silently finds nothing and falls back to the default.

Why it matters

The failure is invisible to a developer on an English-locale machine: the default voice is already English and everything sounds correct. It appears only for users elsewhere, and it degrades to gibberish rather than to silence — so it reads as a broken audio feature rather than a missing two-line configuration.

Speech is also usually the accessibility or attention-getting path, which makes it exactly the feature that must not quietly fail for part of the audience.

How to apply

• Always set utterance.lang explicitly to the language of the string.
• Select a matching voice rather than trusting lang alone; some engines honour the tag only when a voice for it exists: voices.filter(v => /^en\b|^en-/i.test(v.lang)), preferring localService.
• Seed the voice list at startup and refresh it: speechSynthesis.addEventListener('voiceschanged', loadVoices).
• Call speechSynthesis.cancel() before speak(); utterances queue rather than replace, so a repeated alert otherwise stacks up a backlog.