You know the word. You used it in an email this morning. Then the meeting turns to you, and it is gone. What comes out instead is shorter, flatter, and less like you than anything you would ever write.
This is not a knowledge problem, and it cannot be fixed by learning more words. It is a retrieval problem, and it has a mechanism.
Producing a sentence in real time is one of the heaviest things your brain does. You are choosing words, holding the grammar of the sentence you are mid-way through, monitoring how it lands, and planning the next one, all at once. All of that runs through working memory, and working memory is famously small (Baddeley, 2003).
In your first language, most of this is automatic and costs you almost nothing. In a second language, far more of it is assembled by hand, so speaking already runs close to the limit before anything else goes wrong.
Under acute stress, the brain shifts control away from the prefrontal cortex, the region doing exactly the assembling described above, toward faster and more reflexive circuits (Arnsten, 2009). Useful when the threat is physical. Ruinous when the task is finding the subjunctive.
Anxiety also spends working memory directly. Anxious thought is not free: monitoring yourself, imagining the listener's judgment, and replaying the last stumble all occupy the same limited system your sentence needs (Eysenck and colleagues, 2007). In second language research specifically, MacIntyre and Gardner (1994) showed that anxiety degrades performance at every stage, taking language in, processing it, and producing it.
The old observation from Yerkes and Dodson (1908) still holds the shape of it: a little activation sharpens performance, but on complex tasks, past a point, more pressure means worse output. Speaking a foreign language in front of people is about as complex as tasks get.
Language is not a neutral skill. Horwitz, Horwitz and Cope (1986) pointed out that almost no other subject implicates your sense of self the way language does, because speaking badly feels like being someone lesser, not just performing badly. That is why the fear runs deeper here than in, say, mathematics. You are not worried about the answer. You are worried about who you sound like.
Krashen (1982) called the consequence the affective filter: when anxiety is up, language that you fully possess simply does not come through.
If the words vanish because retrieval is competing with fear for the same narrow channel, then the fix is not more study. It is changing the conditions of practice: speaking regularly in a setting where nothing is scoring you and nobody is judging, until producing the language stops being an event. What runs on habit makes almost no demand on working memory, and what makes no demand on working memory has nothing for stress to take.
That principle is the foundation UnMute is built on. The room never judges, and the practice continues until the speaking is boring. Boring doesn't panic.