Free, Offline SRT Subtitles With OpenAI Whisper: A 4-Step CLI Workflow

Описание и применение

Для коротких объясняющих видео автор хочет точных субтитров, но YouTube-аutocaptions портируют имена и термины. Использует локальный开源 Whisper, который пишет .srt за одну команду.

Задача

Short explainer videos need accurate captions, but YouTube's auto-captions mangle names and tech terms; paid caption tools exist but are costly.

Результат

Captions are more accurate than YouTube auto-captions, transcription is fully local after one-time install, no recurring cost, and the same run produces a reusable transcript for descriptions, blog posts, and clip editing.

Рабочий процесс

• Install ffmpeg (Ubuntu/Debian via apt, macOS via brew, Windows via Chocolatey) and install openai-whisper via pip
• Pick a model: start with turbo (optimized large-v3); use tiny.en or base.en for English-only on older hardware; use medium or large for translation
• Generate subtitle file with whisper my-video.mp4 --model turbo --output_format srt --output_dir subs; optionally add --language English, --output_format vtt, or --task translate
• Open the .srt in a text editor and skim it; fix proper nouns, brand names, and numbers before uploading
• Upload the .srt to YouTube Studio via Subtitles > Upload file > With timing
• Optionally reuse the .txt transcript for video descriptions, blog drafts, and clip selection

Доказательства

I make short explainer videos, and for a while I relied on YouTube's auto-captions. They're okay, but they mangle names and tech terms, and I kept wanting a clean transcript I could reuse for descriptions and blog posts.

Источник

Whisper's code and weights are MIT-licensed. You need the internet once, to install it and download a model. After that, transcription is local and your audio never leaves the machine.

Источник

Start with turbo. It's the default, an optimized large-v3, and the accuracy loss is small.

Источник

Before uploading, open the .srt in a text editor and skim it. Whisper is good, but proper nouns, brand names and numbers are where it slips most, and how accurate it is depends on the language and the audio quality. Five minutes of fixes is usually enough.

Источник

If you'd rather script it (batch jobs, a watch folder, a CI step), the Python API is short.

Источник