Kylone Headend · for operators

Captions Module

The "ASR Captions" pre-processing module (package `hwsrnv`) adds a text subtitle page to a channel from two sources: the channel's own DVB bitmap subtitles read with OCR, or the speech on the audio track recognised with a neural model. This note is for the operator who installs it and assigns it to channels.

What it does Requirements Assigning and configuring What to expect Checking it works Known limits

September 2026, MicroCMS v4.

1. What it does

The module sits in the pre-processing tier of a channel. It holds the stream for a short, fixed delay, recognises text, and injects a teletext subtitle page (default 888) into the stream, announced in the PMT with its language. From there the page reaches every tier: the original multicast, the transcoded and stand-in outputs, the set-top boxes and TVs, and the HLS output as a WebVTT rendition when the HLS profile's HLS Text Subtitles option is set to WebVTT from Teletext.

Two sources, chosen per channel:

Text Source	Reads	Delay added	Typical use
DVB subtitles (OCR)	the broadcaster's bitmap subtitle track	1.5 s	channels that carry DVB subtitles but no teletext page: makes them selectable on TVs that only decode teletext subtitles, and available on HLS as text
Speech recognition	the audio track	6 s	channels with no subtitles at all: assistive captions in the spoken language, in English, or both
Auto	DVB subtitles when the service announces them, speech otherwise	1.5 s / 6 s	mixed line-ups where one profile serves many channels

2. Requirements

A GPU: an RTX-class NVIDIA card (Ada or newer) for speech recognition; the lab's GTX 1660 Ti works for tests. OCR runs on the CPU and needs no GPU.
The module package with its language models: models/ggml-*.bin for speech (Small 466 MB, Large v3 Turbo 547 MB, Base 142 MB) and models/tessdata/*.traineddata for OCR (27 languages shipped; eng is the fallback).
One module instance per channel. A stream instance frees a pre-processing slot the way a CAM does; see the stream-instance notes.

3. Assigning and configuring

Create a pre-processing profile from the module and assign it to the channel. The options:

Option	Values	Notes
Text Source	Speech recognition · DVB subtitles (OCR) · Auto	see above
Model	Small (default) · Large v3 Turbo · Base	Speech only. Turbo is the most accurate and needs about twice the GPU time of Small. Base is for smoke tests.
Spoken Language	Detect · English, Italian, Turkish, German, French, Spanish, Portuguese, Dutch, Russian, Arabic, Greek, Polish	Speech only. Setting it avoids mis-detection on multilingual channels.
Captions	In the spoken language · Translated to English · Native page + English on the next page	Speech only. "Both" uses two pages (888 and 889 by default) and doubles the GPU time. Translation is English only.
Acceleration Node	GPU index · Off	Off runs speech on the CPU: only for tests.
Subtitle Track	First announced · a language	OCR only. Which DVB subtitle track to read when the service carries several. Check the channel's languages line in the scanner: `S[eng:10:0xba6]` entries are the DVB subtitle tracks.
Page	888 (default) · 889 · 777 · 778 · 150	The teletext page number. Use a page the broadcaster does not already use.
Page Language in PMT	Same as the source · a language	Overrides the language the page is announced with. With "Both", only the native page; the English page is always `eng`.
Stream Delay	Default · 1.5 · 3 · 5 · 6 · 8 · 10 s	Default is 6 s for speech, 1.5 s for OCR. Longer delays give the speech recogniser more room on a slow GPU.

Make the transcoding profile of the channel copy subtitles (subtitle index copy, tracks all), so the page survives into the transcoded and stand-in tiers.

For HLS text, set in the HLS output profile: HLS Text Subtitles = WebVTT from Teletext, Teletext Page = the page, or All announced pages to publish every page the channel carries.

4. What to expect

OCR: a subtitle appears on the page at the moment the broadcaster shows it and disappears when the broadcaster removes it. Reading a subtitle takes about 20 ms; accuracy on broadcast fonts is high (typically 90 % and above word confidence). Lines are kept as the broadcaster broke them.
Speech: phrases appear as the words are spoken, rolling over the last two rows, and the page clears after a pause. Phrases commit at word boundaries; expect the usual recognition errors on names, numbers and cross-talk. These are assistive captions.
Load: Small takes about 0.8 s of GPU time per 8 s of audio on a GTX 1660 Ti (roughly a tenth of the card per channel), Turbo 1.4 s; "Both" doubles it. RTX cards are several times faster. OCR load is negligible.
Languages on the page: Latin scripts, including Turkish, Polish, Czech, Romanian and the like, are shown with base letters where the teletext Latin set has no accented form. Cyrillic, Greek and Arabic subtitles are read correctly but cannot yet be shown on the teletext page.

5. Checking it works

The channel's pre-processing log (<channel>_pre.log) shows the module's lines: the recogniser opening, pmt: teletext page 888 declared on pid …, then one line per caption (sub … for OCR, cue … for speech, with the language and confidence).
The scanner's languages line for the pre-processed tier lists the page: T[eng:2:888:0xbad].
On a set-top box or TV, open the subtitle menu on the channel and pick the page's language; on HLS, the player's subtitle menu lists the rendition.
The status monitor shows the module's caption count, average recognition time and, for speech, the recognition backlog in seconds. A growing backlog means the GPU is too slow for the model: choose Small, or a longer delay.

6. Known limits

One page language per source; speech translation targets English only.
The speech recogniser needs clean dialogue; music beds and heavy accents lower accuracy.
A DVB subtitle track is read as a whole; region colours and positioning are not carried.
The page is teletext: TVs that decode only DVB bitmap subtitles will not show it. Every hospitality TV platform in current use decodes teletext subtitles.