Dev fine-tuned 1.2B-4B models to clean voice transcripts locally — beats cloud APIs
stephan273 · reddit · 2026-10-28
The developer of voice-to-text app WhisperNow (MIT, open source) fine-tuned his own LLMs for the transcript cleanup step — removing fillers, false starts, self-corrections, and formatting lists — running fully locally.
Dataset: 66,557 rows combining public datasets (aawaaz, handy, Disfl-QA) plus synthetic dictations, with cleaned versions generated by a cloud model; four weak spots (long dictations, misheard words, cut-off words, literal correction phrases) were patched with targeted data and retraining.
Models (full fine-tunes on a single rented H100/H200; 100-word dictation on RTX 5070): LFM2.5 1.2B runs in 0.18s using 1.6GB, LFM2.5 2.6B in 0.34s, Qwen3.5 4B in 0.62s — vs 1.4s via Claude Haiku 5.5 / gpt-6-luna APIs.
Eval: a judge model counts factual errors over 4,170 held-out/benchmark rows. Under 25 words all three beat glm-5.3-flash; the 4B beats it at any length and matches max-reasoning glm-5.3 up to 100 words. The off-the-shelf LFM2.5 2.6B makes 4x the errors of the fine-tuned one. Among cloud models, Haiku 5.5 and gpt-6-luna offer the best price-to-performance.
More from coding & agent
- GitHub demos Copilot Code Review Max with stacked PR merge at Universe — kimmonismus · 2026-10-29
- IBM TechXchange: Three Big Challenges of Adopting AI Agents in the Enterprise — craigmullins · 2026-10-29
- Pydantic Author Samuel Colvin on Setting Up Feedback Loops for Code-Writing Models — samuelcolvin · 2026-10-29
- Open-Source Coding Tools Climb T3 Charts: opencode Hits #2 Most Engaged Repo — MeetPatelTech · 2026-10-29
- Project manager says AI agents now automate 70% of the job — here's the full workflow — Theinkling620 · 2026-10-29
- Vercel engineer builds video.md: a motion-to-audio video system for humans and agents — JohnPhamous · 2026-10-29