Five-hour transcript pipeline

An engineer said a full subtitle workflow that once took days can now be pushed through in about five hours with a chain of AI tools and a final human check. (x.com) The pipeline starts with transcription, which turns speech into text with timestamps, then moves to speaker diarization, which is the step that labels who spoke when. The engineer said the stack used Whisper.cpp for the first pass, GPT‑5.4‑mini for cleanup and speaker identification, Gemini for Japanese translation, Claude Code for SubRip subtitle formatting, and DaVinci Resolve for the edit. (x.com) Whisper.cpp is an open-source C and C++ implementation of OpenAI’s Whisper speech-recognition model, and its maintainers describe it as a high-performance way to run transcription locally on devices from Macs to Windows PCs. OpenAI describes Whisper itself as a multilingual speech-recognition model that can transcribe speech, identify languages, and translate speech to text. (github.com, github.com) Speaker diarization is the part that makes a transcript readable in interviews, podcasts, and meetings, because it splits one long block of text into turns by Speaker 1, Speaker 2, and so on. That step is still often stitched together from separate tools rather than handled cleanly in one pass, as current Whisper.cpp workflows and third-party guides show. (picovoice.ai, deepwiki.com) Translation and subtitle export are the next bottlenecks, especially for Japanese, where line breaks, punctuation, and reading speed can make a subtitle file usable or unusable. Google says its translation products now span 189 languages, and Anthropic describes Claude Code as a coding agent that can edit files and run terminal workflows, which fits the engineer’s use of it to turn cleaned text into properly structured subtitle files. (cloud.google.com, anthropic.com) Those subtitle files matter because DaVinci Resolve can import SubRip, or SRT, captions directly into an edit timeline, letting editors review timing and fix mistakes inside the video project instead of rebuilding captions by hand. Blackmagic Design’s current Resolve support notes also show active subtitle updates, including improved subtitle kerning in Resolve Studio 20.3.2. (blackmagicdesign.com, gotranscript.com) The engineer did not present the workflow as fully automatic. In the post, he said the speed gain was real but some tasks were still inconsistent enough that a person had to review the output before delivery. (x.com) That caveat matches the way these systems are being used in production in 2026: one model handles raw listening, another cleans structure, another translates, and another formats files for software that editors already use. The result is less typing and fewer manual passes, but not a no-touch pipeline. (github.com, openai.com, anthropic.com, blackmagicdesign.com) The five-hour claim lands less as a promise that subtitle work is solved than as a snapshot of how creators are assembling model-by-model workflows around old post-production tools. The machine does the first draft at speed; the editor still decides what is safe to publish. (x.com)