Best Models for Transcript Cleanup Guide | Generated by AI
Great use case. For long-form transcript cleanup with strict formatting and “no extra chatter”, here’s what works best right now on OpenRouter, by scenario.
Fast + cheap workhorse (EN/zh, long files): DeepSeek-V3.1 Strong at instruction-following, handles 128k context, and you can toggle “reasoning” off for speed or on when the text is messy. Good bilingual performance for filler-word removal and consistent speaker tags. (DeepSeek API Docs, OpenRouter)
Chinese-heavy interviews & colloquialisms: Kimi K2 Instruct Moonshot’s K2 (MoE) is particularly fluent with Chinese slang and style; great for zh-first transcripts while keeping technical nouns intact. (OpenRouter)
Highest compliance on editing instructions: Claude Sonnet (3.7/4) Anthropic’s Sonnet line is excellent at “output only the refined text, no meta,” and tends to be conservative about meaning changes—ideal for your step-list constraints. Use Sonnet 4 if available; 3.7 also performs well. (OpenRouter)
Ultra-long sessions or one-shot whole-project passes: GPT-5 When you want to push very large contexts and keep hallucinations low, GPT-5 is the safest bet among frontier models on OpenRouter (listed with a very large context; strong reasoning and editing). Use for marathon transcripts or final “polish” passes. (OpenRouter)
Also strong, but watch cost profiles: Gemini 2.5 Pro Very capable at reasoning and long-context editing. It’s solid for refinement, but mind pricing/quotas depending on your provider route. (OpenRouter)
A practical routing recipe (no tables)
- ≤128k tokens, EN/zh mix, speed matters: DeepSeek-V3.1 (non-thinking). Flip to thinking only when paragraphs are chaotic. (DeepSeek API Docs)
- Primarily Chinese transcripts: Kimi K2 Instruct. (OpenRouter)
- Strict “editor” behavior (no commentary), legal/finance tone: Claude Sonnet. (OpenRouter)
- Gigantic files or final single-pass polish: GPT-5. (OpenRouter)
Quick code tweaks to make this robust
- Make model selectable: add
--modelCLI flag; default todeepseek-v3.1. - Auto-route by size/language: detect token count and zh ratio; pick K2 for zh-heavy, GPT-5 for huge inputs, Claude for “strict edit” mode.
- Reasoning toggle (DeepSeek): add a param to set
reasoning.enabledin your OpenRouter call for hard passages. (DeepSeek API Docs) - Fix small bugs: that stray
print(".1f"); and widen speaker-label regex to allowA：(full-width colon) and no-space variants. - Guardrails: if length delta >10%, auto-retry with a “don’t delete content” reminder; if no
A:/B:found, auto-retry with a “label speakers” nudge.
If you want, I can rewrite your script to add the --model flag and smart routing logic right now.