Home › Guides

The translation glossary (KNP): the one document that stops your character's name changing spelling in episode five

By the Fonti Studio post-production team · 28 October 2025

Here is a failure mode every distributor who has shipped a series recognises. Episode one spells the lead character Shauna. Episode five spells her Shana. Episode seven invents a third option. Nobody was careless; three different linguists worked three different episodes, each made a reasonable call, and no document existed to make the call for them.

The document that prevents it is a glossary, and on a platform delivery schedule you will meet it under a specific name: the KNP, or Key Names and Phrases list. It is one of the least glamorous items in a localization package and one of the most consequential, because it is the only thing standing between a multi-language, multi-episode deliverable and terminology that quietly contradicts itself.

What a KNP actually is

A KNP is a controlled list of every term in a title that must be rendered consistently, with its approved form in the source language and in each target language. Netflix publishes a Key Names and Phrases template for exactly this purpose, and the fact that a platform maintains tooling for it tells you how routinely the problem bites.

What goes on the list:

• Character names, including nicknames, diminutives, and the forms different characters use for the same person.
• Places: cities, regions, invented locations, buildings that recur.
• Organisations: agencies, companies, crews, factions.
• Invented terminology: anything the script capitalises that does not exist outside the film. Genre titles are dense with these.
• Recurring phrases: catchphrases, mottos, chants, anything an audience will hear repeatedly and expect to see rendered the same way.
• Titles and forms of address, which carry the register decisions that a translator would otherwise re-litigate every episode.

And, critically, a column for the decision: does this term get translated, transliterated, or left exactly as it is? That single choice, made once and written down, eliminates most terminology drift before it starts.

Where the terms come from

The source of truth depends on what you have.

With a CCSL, the work is extraction rather than invention: the character map already names every speaker in canonical spelling, and the spotting list already shows the terms as authored. Everything the glossary needs is written down by someone who watched the film frame by frame.

With a source-language subtitle instead, the terms come from the subtitle's own spelling, which is usually reliable and occasionally not. On one reference title our verification pass found the supplied master itself misspelled three cast names against the production's own spelling. That is worth knowing before you propagate it into six languages, which is the second reason to build the glossary as an explicit, reviewed step rather than a side effect of translating.

With neither document, the terms have to be recovered from the audio, and proper nouns are exactly where automatic transcription is weakest, because an invented term carries no linguistic redundancy for a model to lean on. We treat that case as a flagged risk in the brief rather than a quiet assumption.

Why models drift, and why a prompt is not a glossary

It is tempting to think a modern translation model simply remembers. It does not, in the way this problem requires.

Translation runs in chunks. Each chunk is a separate generation, sampled independently, and a name that appeared forty minutes of runtime earlier is not reliably in view. Where a name has more than one plausible rendering, and most transliterated names do, the model will pick differently on different occasions, correctly each time, inconsistently across the file. Across a season with separate runs per episode, the drift compounds. Across languages, it multiplies again, because each language makes its own transliteration choices.

The fix is not "remind the model," it is constraint. The locked glossary is injected into every translation call for every chunk, and the QA pass afterwards checks the output against it: if a locked term appears in a form other than the approved one, that is an error to fix, not a stylistic variation to consider. The glossary also has to be handled with care by the automated checks around it, since a cue that reads "Michael Rivers." is entirely composed of proper nouns and will look untranslated to a naive validator forever unless the checker knows the glossary.

On our reference delivery, The Last Kumite (2024) into French and Dutch, that lock held 47 cast, place, and term entries with zero drift across 1,212 cues per language, and the same list fed the dub's casting roster so the name in the subtitle and the name attached to the cloned voice were byte-identical.

The review step, which is the point

A glossary that a machine builds and nobody reads is a liability with a false sense of security. Extraction gets the candidate list; a human decides.

The decisions that need a person are the ones with no mechanical answer. Does this nickname translate, or is it part of the character's identity? Does the invented faction name get localised, and if so does the localised form still work in the mouth of a dubbed actor? Is this street name a real place the audience will recognise or set dressing? Does the title card carry an official localised title, or does it stay in the original?

This is why we build the glossary as its own explicit step with its own output file, review it with the client once, and only then start the translation loop. Ten minutes on a name list at the start is worth more than any amount of correction later, because correction after the fact means re-running the affected cues in every language and re-verifying them.

What to ask your vendor

Three questions separate a vendor with a real glossary discipline from one without.

Ask when the glossary is built. If the answer is "as we go," terms are being decided per chunk, which is the drift condition.

Ask whether you get to see it. If there is no artifact to review before translation starts, the editorial decisions are being made without you, and you will meet them at delivery.

Ask how it is enforced. "The translators have it" is a policy. A validator that fails a delivery when a locked term appears in an unapproved form is a mechanism. For a series in particular, ask whether the glossary is locked once per show or rebuilt per episode, because the whole point is that it spans episodes.

Common questions

What is a KNP in localization?

A Key Names and Phrases glossary: the controlled list of character names, places, organisations, invented terms and recurring phrases for a title, with the approved rendering in each target language. Platforms including Netflix provide templates for it, and it exists so that everyone touching the title uses identical strings.

Why do subtitle translations spell names inconsistently?

Because translation happens in chunks, each generated independently, and any name with more than one plausible spelling or transliteration can be rendered differently on different occasions. Across separate episodes and separate languages the variation multiplies. A glossary locked before translation and enforced by a validator afterwards is the structural fix.

Who decides whether a name gets translated?

The rightsholder, in review with the vendor. Some names are identity and stay as authored; some invented terms need localising to land; a few (an official localised film title, for instance) are brand decisions that only the distributor can make. The vendor's job is to surface every one of those choices before translation, not to make them silently.

Do I need a glossary for a single feature, or only for series?

Both, for different reasons. A series needs it to stay consistent across episodes and linguists. A feature needs it to stay consistent across languages and, if you also commission a dub, to keep the subtitle spelling and the dub casting roster in agreement.

The short version

The glossary is the cheapest quality control in a localization package and the one most often skipped, because its value is invisible when it works. Build it first, review it once, enforce it mechanically, and the class of defect that makes a series look careless simply cannot occur.

Fonti Studio builds a locked glossary per title before a line is translated, sends it to you for review, constrains every translation pass to it, and validates the delivery against it. It is included in the flat price, not a line item. If you have a series where episode consistency is the thing keeping you up, email us the brief.

Fonti Studio is a subtitle and dubbing studio run by post-production engineers, for film distributors and sales agents. €400 per language for feature subtitles, €100 to €200 per language per episode, glossary locked once per show. Flat. EUR. Free preview before you pay.