Home › Guides

Netflix subtitle requirements: what indie distributors get wrong about the Branded Delivery spec

By the Fonti Studio post-production team · 2 September 2025

In short: most rejected Netflix subtitle deliveries fail on measurable numbers, not on translation: reading speed over 20 characters per second, cues crossing shot changes, lines over 42 characters, gaps under two frames, format non-conformance. The Branded Delivery spec publishes every limit, so a file can be checked against all of them before submission. That is what gets a feature through QC on the first pass.

Most rejected subtitle deliveries don't fail on translation. They fail on a number.

The usual culprits:

• Reading speed: above 20 characters per second.
• Shot changes: a cue that runs across a shot change.
• Line length: a line longer than 42 characters.
• Timing gaps: a gap under two frames.
• Format conformance: a non-compliant character in a TTML attribute.

The QC tool catches it, the file bounces, and the distributor loses a slot in the platform's release calendar, a window that may not open again for a quarter.

This piece is the working guide for indie distributors and sales agents who need their localized masters to pass Netflix Branded Delivery on the first submission. Not the Originals spec (a separate, stricter regime that very few indies will need to hit). The Branded spec, the one that governs everything Netflix acquires, and the five places we routinely see AI-captioned files fail it.

What "Branded" means

Netflix runs two parallel delivery specs.

Netflix Originals is the spec for content Netflix produces or co-produces. It enforces a deeper set of constraints, including the notorious dialog stem identity rule, which requires the 5.1 dialogue stem summed with M&E to bit-reconstruct the full 5.1 mix. That's a constraint AI-separation pipelines can't meet; separation is lossy by definition.

Netflix Branded Delivery is the spec for packages that carry the Netflix presentation credit: Originals, and licensed titles that Netflix brands as its own in some territories. A plain licensed acquisition with no Netflix card falls under the Non-Branded specification instead. On the subtitle side the two are close relatives (same formats, same timing rules, same separate forced-narrative and SDH files), and the Branded document is the fuller, public reference, so that is the one this piece quotes; its subtitle numbers are the ones a licensed delivery is measured against too. The text in the file has to read; the timing has to track the cut; the formats have to parse. (Who actually makes the subtitles on a licensed title, which spec applies, and what the package has to contain beyond the numbers below, is covered in Netflix subtitling: how subtitles get onto Netflix.)

Netflix's Localization, Accessibility and Dubbing Branded Delivery Specifications is the public reference. Every claim in this piece is grounded in it.

The five places AI-captioned files fail

1. Reading speed (CPS).

• The limit: The Branded spec caps reading speed at 20 characters per second for adult content. A line that reads at 22 CPS fails QC, even if every word is correctly translated.
• The AI failure mode: a model that translates faithfully will often produce a French or German cue 30% longer than the English source, because the target languages re-distribute the same meaning across more words. If the cue duration doesn't move with it, the CPS spikes. The Whisper-large-v3 + ad-hoc translation stacks we see in the field don't optimise CPS at all, they emit whatever the model produced and let the QC tool reject it.
• The fix: time can be extended (into idle space before the next cue) to lower CPS, but the cue start always has to track the speech onset. You can't fabricate timing; you can only recover it from the silence either side of speech. Our pipeline does this as a post-translation optimisation, and tracks it as a hard QC gate before the file leaves the building. On the Kumite reference run, this kept 1,212 cues across French and Dutch within the 20 CPS ceiling.

2. Shot-change misalignment.

• The rule: The Branded spec doesn't name a frame tolerance, but the platform's QC vendors increasingly catch subtitles that flash in and out across a cut. A cue that ends 3 frames after a shot change reads as a flash to the eye; a cue that starts 5 frames before reads as late. Editors notice both.
• The AI failure mode: a model that segments on linguistic pauses will land cue boundaries exactly at shot changes (because dialogue often pauses around cuts). It's the worst possible alignment.
• The fix: detect the shot list before authoring cues, and apply a 7-frame red zone around each cut: a cue either lands entirely within a shot or extends cleanly across one. Idle space around speech onsets is the budget; the cuts are constraints. Our pipeline does shot detection on the picture and treats it as authoritative.

3. Character-count and line-count overrun.

• The limit: The Branded spec caps lines at 42 characters and cues at 2 lines. A 3-line cue is an automatic reject. A 43-character line is too.
• The AI failure mode: This is a translation problem, not a timing problem. A French re-ordering of an English sentence will sometimes produce a line that fits the cue but overflows the screen width. The model doesn't know.
• The fix: enforce line-length as part of the translation pass, not as a post-hoc reformatter. The model should be told the character budget per cue, and re-roll lines that overflow. Re-splitting a long line into two short cues changes timing, better to constrain the line at generation time.

4. Format conformance. TTML attributes and EBU-TT-D structure.

• The requirement: A correctly translated, correctly timed SRT can still bounce in TTML. Netflix's TTML1 delivery (and the IMSC 1.1 profile it uses for Japanese) is strict on namespace declarations, element ordering, and the handful of style attributes that aren't universally supported. EBU-TT-D (Tech 3350) is the European broadcaster equivalent and has its own quirks (region declarations, metadata blocks, line-height defaults).
• The AI failure mode: an SRT-first pipeline that converts to TTML at the end will often emit a TTML file that parses but doesn't conform to the platform's stricter sub-profile. Conformant TTML is not "an SRT with angle brackets."
• The fix: author in a structured intermediate that's broadcast-spec from the start, and let the deliverable formats render from it deterministically. Our pipeline emits SRT, TTML, IMSC 1.1, EBU-TT-D and CEA-608/708 from the same internal cue representation, and runs each through a per-format conformance validator before it ships.

5. Glossary drift across episodes.

• The risk: The Branded spec doesn't enforce glossary lock, but every distributor who has shipped a multi-episode series knows what happens when episode 3 spells a character name Shauna and episode 5 spells her Shana. The platform's editorial QC will flag it; the consumer will mock it on social.
• The AI failure mode: an episode-by-episode translation pass with no shared glossary. Even with the same model, sampling will drift.
• The fix: extract the glossary from the source CCSL (or, if no CCSL, from the first episode's source SRT) and lock it across the whole deliverable set. Every name, place, term, and capitalisation is canonical; the translation pass is constrained to those exact strings. On our pipeline, that glossary is also the input to the dub side's voice-cast roster, so the name in the subtitle and the name on the casting list are bit-identical.

What we keep that AI captioners drop

Two more things, less visible to the platform QC but obvious to the buyer the first time they spot-check the file:

Register. Tu or vous. Du or Sie. The model defaults to the neutral, polite form. The drama requires the casual one. A faithful translation that misses the register is a failure of meaning, not of text. Our pipeline runs a cross-model verification pass after the first translation specifically to catch register regressions.

Idiom. "On the way to losing it" doesn't translate as "en allant le perdre" by accident. It does so because the translator was constrained to fit the cue budget and preserve the philosophical shape of the line. Most AI captioners optimise neither.

What this looks like on the Kumite reference

The Last Kumite (2024), 101 minutes, English source, delivered in French and Dutch:

• 1,212 cues per language
• Every cue ≤ 20 CPS
• Every cue ≤ 42 chars/line, ≤ 2 lines
• Every cue ≥ 5/6 second minimum duration, ≥ 2-frame gap
• Zero shot-change overlaps
• Glossary-locked: 47 cast/place/term entries across both languages
• Zero errors against Netflix Branded Delivery spec, verified by automated conformance check before the file left the building

That's not a quality target. That's the floor. Every Fonti Studio delivery ships against it.

Fonti Studio is a subtitle and dubbing studio run by post-production engineers, for film distributors and sales agents. Broadcast-grade output validated on feature-length deliveries. €400 per language for subtitles. €3,500 per language for dubs. Flat. EUR.