OpenSubtitles

OpenSubtitles, as packaged within the OPUS collection of open parallel corpora, is a vast pile of film and television subtitles aligned across languages. The current version spans 60 languages and around 2.6 billion sentences, offered both as monolingual text per language and as sentence-aligned pairs between languages. Because subtitles are dialogue, the text captures how people actually speak: questions, slang, interjections, and the short conversational turns you rarely find in formal writing.

You download it directly from the OPUS site, picking a language or language pair and a format. Moses format gives you plain aligned text files, TMX is the translation-memory format, and the XML version preserves subtitle timing and structure. It is also mirrored on Hugging Face for streaming. For RAG, the aligned pairs suit translation and cross-lingual retrieval, while the monolingual text works well as conversational training or grounding data.

It shines for machine translation, for dialogue and chat-style systems that need colloquial language, and for low-resource language pairs where cleaner corpora simply do not exist. If your application has to sound like natural speech rather than an encyclopaedia, this is a strong ingredient.

Alignment is derived from subtitle timing, which makes it noisy: expect misalignments, out-of-order lines, and the occasional mismatch, plus profanity and plot spoilers that come with the source material. On licensing, the corpus compilation is free for research, but the underlying subtitles were written by many uploaders under varied and often unclear terms, so treat commercial use as restricted and keep things to prototypes unless you have cleared the rights.

It sits within the wider OPUS family, so if you need cleaner or larger translation data, ParaCrawl and CCMatrix are worth comparing. OpenSubtitles wins on conversational tone rather than pristine alignment.