Quick question: while I love the offline aspect, how does this handle spelling in relation to context? Is that via a ruleset, or is there some intelligence that learns user speaking patterns and common subjects?
Good question! The short answer is, neither. There are no hardcoded rules, but the app also doesn't actively learn your personal speaking patterns over time.
All the context-awareness comes straight from the pre-trained Whisper model. Since it's a transformer network, it looks at the entire sentence context rather than translating word-by-word. For example, if you dictate a sentence about coding, it naturally knows to capitalize "Rust" and "Python" instead of writing about rusty metal and snakes.
I deliberately kept the model static. Trying to fine-tune it locally on the fly would mean I'd have to store your voice data (which kills the 100% privacy promise).
That being said, adding a custom dictionary feature, so you can feed it highly specific industry jargon right before you speak, is at the very top of my to-do list!
Let me know how it handles your vocabulary if you give the trial a spin.
Grüße as Kanada!!
Quick question: while I love the offline aspect, how does this handle spelling in relation to context? Is that via a ruleset, or is there some intelligence that learns user speaking patterns and common subjects?
Grüße zurück nach Kanada!
Good question! The short answer is, neither. There are no hardcoded rules, but the app also doesn't actively learn your personal speaking patterns over time.
All the context-awareness comes straight from the pre-trained Whisper model. Since it's a transformer network, it looks at the entire sentence context rather than translating word-by-word. For example, if you dictate a sentence about coding, it naturally knows to capitalize "Rust" and "Python" instead of writing about rusty metal and snakes.
I deliberately kept the model static. Trying to fine-tune it locally on the fly would mean I'd have to store your voice data (which kills the 100% privacy promise).
That being said, adding a custom dictionary feature, so you can feed it highly specific industry jargon right before you speak, is at the very top of my to-do list!
Let me know how it handles your vocabulary if you give the trial a spin.