Subtitles Generator

Subtitles Generator turns a video into subtitles with one command: subtitle video.mp4 --to en,fr detects the spoken language, transcribes it with Whisper, times every word to the audio, and writes .srt files in the original language and in however many other languages you ask for.

It’s a spin-off of Video Translator , stripped down to just the subtitle half of that pipeline: no voice cloning or dubbing, just accurate, well-timed text in as many languages as needed.

Per-word timing, not per-sentence

Whisper’s own timestamps are per-segment, which tends to produce subtitle lines that start too early or linger after the speaker has moved on. Each word is aligned to the audio separately, so a line’s start and end match the speech, and long sentences split between words rather than wherever Whisper happened to end a segment.

A glossary so names survive transcription and translation

Whisper regularly mangles names and place names, and a plain translation mangles them again. A JSON glossary maps a source term to how it should read in the target language; its source-language spellings are fed to Whisper as a prompt so it transcribes them correctly in the first place (for Japanese and Chinese, this also tends to improve punctuation and line breaks), and the same glossary keeps every target language’s translator consistent about how it renders those names afterward.

Two translators, different trade-offs

gpt (OpenAI, or a local model through Ollama) uses the glossary and a one-line description of the video to translate naturally, with neighbouring lines as context. google translates literally, line by line, without either. GPT costs more per language but reads better; Google is cheaper and needs a Google Cloud service account instead of an OpenAI key.

Resumable by design

Every finished step is recorded per video, so re-running the same command to add another language only runs translation, not transcription and alignment again. An interrupted run picks up where it left off instead of starting over.

The repo README has the full setup and usage instructions, including an optional --separate step that strips music and background noise before transcription.