Practical experience: Fine-tuning a 0.8B model for local ASR post-processing

MoodOdd9657 · reddit · 2026-08-18

The author, while building the open-source dictation app SpeakoFlow, identified the challenge of post-processing ASR transcripts: correcting self-corrections, structuring rambling speech, and interpreting instructions without over-editing. Experiments showed that 0.8B models improve significantly with detailed prompts, suggesting fine-tuning could outperform larger general models. The plan involves fine-tuning Qwen3.5 0.8B using QLoRA with a mix of real and synthetic data, then deploying via llama.cpp. The author also discussed load time differences between Apple Silicon and discrete GPUs and sought advice on base model selection.

Original post →

More from coding & agent

coding & agent channel →