Improve caption quality
Improve caption quality for your live stream
There are a number of things you can do to improve live caption quality.
Increase the HLS segment size in Brightcove
Increasing the segment size of your HLS stream will allow CaptionHub to improve the caption quality. The recommended segment size is 10 seconds (minimum 6 seconds).
- In Brightcove, go to Ingest Profiles and choose the profile that you use for your live streams.
- If you use a default Brightcove profile (like “Premium Live HD 1080p - Tier 2”), click to duplicate the profile
- For each rendition, update the
segment_secondsvalue to 10:
- Save the profile
- Next time you create a live stream, select the new profile
Use custom dictionaries
Custom dictionaries bias the speech recognition towards certain words. In a live context, it is highly recommended that you configure a custom dictionary with speaker names, product names and other specific words that speech recognition might not pick up automatically.
Translations - Use termbases
Termbases are similar to custom dictionaries, but influence how words are translated. Use this to specify specific translations, or to avoid translating brand and product names.
Translations - Enable Live brevity / profanity control
Brevity: Amazon reference for brevity When translating between languages, the resulting translation can sometimes be longer (in character count) than desired. This can pose a problem in certain contexts (such as captions, subtitles, headlines, or form fields) where space for extra characters is limited.
You can turn on the brevity setting when you run real-time text translations with Amazon Translate. Brevity reduces the length of the translation output for most translations (compared to the translation output without brevity).
Profanity: Amazon reference for profanity filter When you run translations with Amazon Translate, you can enable the profanity setting to mask profane words and phrases in your translation output.
To mask profane words and phrases, Amazon Translate replaces them with the grawlix string "?$#@$". This 5-character sequence is used for each profane word or phrase, regardless of the length or number of words.
Amazon Translate does not mask profanity in translation requests where the source language and target language are the same.
In some cases, a profane word in the source input might naturally become inoffensive in the translated output. In such cases, no masking is applied.
Amazon Translate detects each profane word or phrase literally, not contextually. This means that it might mask a profane word even if it's inoffensive in context. For example, if Amazon Translate detected "jerk" as a profane word, then it would write the phrase "jerk chicken" as "?$#@$ chicken", even though "jerk chicken" is inoffensive. (Here, "jerk" is used as an example only. Amazon Translate does not detect that word as profanity.)