Fact · Dictation on Windows
On NPU-equipped devices, Phi Silica uses speculative decoding, with a smaller draft model, to speed up text generation.
Checked against its source on October 16, 2026
In the source's own words
“employs a technique called speculative decoding to accelerate text generation using a smaller draft model”
Phi Silica in the Windows App SDK, Microsoft Learn, 2026, Introduction, paragraph on how the model runs.
Open the original- What it measures
- Speculative decoding has a smaller model propose several tokens that the main model then checks in parallel, which Microsoft describes as the way the NPU speeds up generation.
- What it does not show
- Microsoft gives no speed figure in this passage, so the page cannot show how much faster the NPU makes fluid dictation. It describes text generation in general, not the dictation cleanup.