A 14-year-old trained a single model to replace the Whisper+LLM dictation pipeline: finished text in ~0.5s on 4GB VRAM

Mehmet_dedeler · reddit · 2026-11-14

A 14-year-old developer recreated Wispr Flow-style dictation by merging the "Whisper transcription + LLM cleanup" pipeline into one model. The two-model pipeline survives only at training time: real unscripted speech from openly licensed datasets is paired with LLM-written finished text (clips that dropped real content are discarded), then Qwen3-ASR 1.7B (Apache-2.0) is fine-tuned to go straight from audio to finished text.

Results: the two-model setup took 2.5–36s per dictation on an RTX 3050 Ti (4GB), with models swapping in and out of VRAM and text generated twice; the single model takes 0.5–1s, is a 2.2GB Q8 file using 2.5GB VRAM, and runs in llama.cpp. Quality also improves — a cleanup LLM can't hear vocal cues like restarts, while the single model keeps what's actually meant. Shipped as the free open-source app Typd for Windows/NVIDIA and Apple Silicon Macs.

Original post →

A 14-year-old trained a single model to replace the Whisper+LLM dictation pipeline: finished text in ~0.5s on 4GB VRAM

More from Apps