ADENA Hack: How to change language models on the KS-CC1.

The 2.1.0.4 firmware update for KS-CC1 introduced a number of important changes, including language models: it introduced OpenAI’s Speech to Text API and the V2 update for Google Speech to Text API.

Google and OpenAI models have significant differences: the kind and volume of data they were trained on, how they work with audio fragments, which languages they support, their unique features, and more. There are many detailed head-to-head comparisons between these language models available online, but to keep it short, here are the main differences and their practical effects:

	Google API	OpenAI API
Training data size and origin	Millions of hours of multilingual, self-supervised data from various sources.	680,000 hours of web-sourced, semi-supervised data. At least ⅓ of it is English.
Confidence threshold (similarity percentage required for a captioned word to be displayed)	High	Low
Audio recognition method	Comparison with pronunciation database, prioritising confidence threshold.	Analysis and prediction, prioritising transcription completeness.
Practical effects of these differences	Potentially higher accuracy for transcribed text, but with statistically more words missing from transcription.	Statistically more complete transcriptions, but with potentially higher error rate or “hallucinated” words.

Users are able to freely switch between these models using the “Administrator” web page of KS-CC1. To do that, type the IP address of the station in a browser, open “Administrator” → “Live Transcription” → “Live Caption settings” and select the model you would like to try under “Speech to Text”. The model is changed instantly, and a system restart is not required. Thus, you can try which model works best in your scenario at any time.

Importantly, both models rely on the same translation and transcription licencing mechanism, so there is no need to purchase different licence keys to use another language model.

While the list of supported languages keeps growing, the new models do not yet support the same number of languages as version 1 of Google Speech to Text API. You can check the list of supported languages for both models on version 2.1.0.4 here. If you are updating from 2.0.0.18 or earlier version to 2.1.0.4., you will need to reset the station to factory default settings afterwards.

If a language you need is not supported yet, you can install the 2.0.0.18 firmware version on your station to use the previous API version. You can check the list of supported languages on that version here. Also, note that you will need to reset the station to factory default settings afterwards.

There are more features to highlight in the 2.1.0.4 update, so stay tuned for our next post! As always, if you would like to know more about AREC devices or see them in action, contact us at www.a-dena.com, and we would be happy to assist you.