Translate subtitles
to Polish

How it works for Polish

Upload an English .srt or .vtt file, or a season as a zip. Cues travel to the model in batches of 32 so a sentence that spans two cues is translated as one thought. Every Polish cue then passes thirteen structural checks; whatever fails goes back with the exact complaint attached, and an editor pass sweeps the whole file for register and idiom. You download the subtitles and a quality report, and the upload is deleted.

What the Polish profile checks

• Sound labels are nouns: [ŚMIECH], [MUZYKA], [STRZAŁ]. A clause like [ONA SIĘ ŚMIEJE] needs aspect, gender and number to agree, and a table that gets one wrong emits broken text with no model to catch it. A noun has nothing to get wrong.
• „…” quotation marks by instruction; 46 Polish function words required, 26 known intrusions blocked.
• Structurally among the cleanest outputs measured (16 issues on four episodes) but the profile is a quarter the size of the tuned ones, and the chrF++ of 44.90 reflects that. Labelled beta on the site for that reason.

Sound labels, from the table

Labels for the deaf and hard of hearing are resolved from a fixed Polishtable before the model sees the text, so they are never invented. Three entries, exactly as the table renders them:

[LAUGHS]

[ŚMIECH]

[DOOR CLOSES]

[ZAMKNIĘCIE DRZWI]

[PHONE RINGS]

[DZWONEK TELEFONU]

Is Polish measured?

Polish has a measured baseline: 44.90 chrF++ against a professional human translation, over 2,012 aligned units. That figure is what every change to the pipeline has to beat for Polish. It is not comparable to another language's figure; here is why.

Price and time

One credit translates about 20,000 characters of dialogue, roughly a 25-minute episode; a feature film is about three. Credits are $1.25 each in packs, the first episode is free, and there is no subscription. An episode takes a few minutes; you get an email when it is ready.

Reading

• Why sound labels come from a table, not from the model