Show HN: Teach a speech model (en) a name by saying it once, on your Mac
I wanted to see if a speech model could spell a name better if it heard it, not just read it. 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
Read original source
Source record
Recorded metrics
points · 2 votes
2026-11-05 21:15:31.861Z
cumulative since publication
Research analysis
Research claim
Source account: I wanted to see if a speech model could spell a name better if it heard it, not just read it. 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
Evaluation record
Source votes: 2 votes (Nov 5, 2026 UTC; cumulative since publication) Recent records to compare: “Memorilla: Latent Semantic Memory for LLMs” (Hacker News, Nov 6, 2026); “Show HN: Ru-Jev – an open model that scores caller-supplied options on CPU (MIT)” (Show HN, Nov 6, 2026). Shared themes are navigation hints; they have not been verified as the same need or independent corroboration.
What the result supports
Interpret results only for the reported model version, task, data and inference budget. Community votes, downloads or a benchmark position do not establish reliability in a different workflow.
Validity gaps
Check sample selection, baseline configuration, uncertainty, contamination risk, negative results and access to code or weights. No reproduction by WeWorkBuddy is recorded here.
Verification plan
Read the paper or model card and its evaluation appendix. Reproduce one bounded result against the stated baseline, then test a representative user task and record where the conclusion fails.