Show HN: Teach a speech model (en) a name by saying it once, on your Mac
Most STT tools handle custom words as text (keyterm list, fixing transcript after with llm, etc).
So I trained an adapter to try it, specifically a LoRA on Qwen3-ASR 1.7b (see benchmark). This LoRA transcribes words only after it hears a recording of each of your custom words.
To make this model easier to use in day-to-day life, we also built Detail Keeper. It's a Mac app where you can put this adapter to use by specifying difficult-to-transcribe-correctly words and then having them picked up correctly in your recordings.
We've found this helpful in two specific cases: 1. Words STT has never heard: invented characters and places, codenames, unusual surnames. 2. Your spelling of a word it does know: Kathryn and not Catherine, Jon and not John.
I wasn't able to achieve both good recognition and recognizing many words (1k+) with the model alone.
Detail Keeper first narrows each passage to the likeliest taught names by sound. Then the adapter writes the passage, and the app checks it against a plain transcript.
On our open benchmark (audiobooks, 1,100 name mentions) it spells taught names as well as Eleven Labs Scribe with keyterms: 89.6% vs 88.2%, no measurable difference. The only downside is that it overwrites a few more untaught names and changes slightly more ordinary words.
Everything, including where we lose, is at detailkeeper.com/benchmark.
Note: It's free, local and nothing leaves your Mac, but it does require an Apple M series+ mac and OS 26+ currently. Also worth noting that it downloads about 3.6GB of models and only works in English right now.
If you're interested, you can also find the adapter itself on HF: huggingface.co/virgil-coach/own-names-qwen3-asr