Best Practice for Customers When Adding Materials in the Platform Follow
Adding preparation material to your order
Verbit’s Automatic Speech Recognition (ASR) technology uses a series of models consisting of an acoustic model (AM) which identifies sounds, followed by a phonetic or pronunciation model (PM) which produces words from a series of sounds, and a language model (LM) which attaches those words into sequences in order to create meaning. While Verbit’s ASR is one of the most reliable Speech-to-Text (STT) systems available in the transcription market; it is the use of area-specific models and Verbit’s “human layer” which are responsible for Verbit’s superior output.
Both the area-specific models and Verbit’s “human layer” rely on the transfer of preparation material to build “the glossary” in order to remain relevant.
Uploading preparation material helps with the output
Verbit clients have an important role to play in the personalization process.
By sharing idiomatic and distinctive terminology pertaining to the transcription, meeting, or event; Verbit users can train the Automatic Speech Recognition to detect and understand topics and niche terms which in return, deliver users with superior accuracy and increases the chance to receive a customized meeting or event experience.
To accurately transcribe singular terms, Verbit’s Automatic Speech Recognition requires users to upload a list of people's names, brands, places as well as specific professional terminology ahead of time.
Take for example Soccer players' names: Lionel Messi or Neymar. If these names are not present in the dictionary, a sentence such as: “Messi with the ball” may be recognized as “Mess with the ball” which has a totally different meaning and is wrong of course. Instead of relying on the recognition model vocabulary and probabilities alone, the customized dictionary allows for a prepared list of words such as people, places, and professional terms to become recognized.
In addition, sharing short paragraphs of content provides a context of reference known as “terminology” around the topics being covered which in return, increases the probabilities for the language model to build more relevant sentences, especially around complex topics.
Take for example the term Corona. Prior to Covid days, Corona was mainly known as a beer, so the probabilities for sequences such as “Corona beer” or “Let’s have Corona” were high, whereas the probability of “Corona Virus” was low or even zero. Adding texts where the combination of words “Coronavirus” or “I came down with Corora” helped the ASR to generate the correct transcription when discussing Covid issues. Providing text increases the probabilities of positive recognition and disambiguation for the language model. A unigram will generate the same acoustic sounds but providing context will help differentiate words to use
If you are using Verbit’s auto solution (ASR), you can pre-upload course materials, keywords, and names, which will be used by Verbit’s language models to pre-train the AI to detect these terms quickly and accurately.
If you are using Verbit’s professional solution (ASR + human layer), Verbit’s transcribers will have access to the same uploaded information and it will help them accurately spell unique terms or names when making corrections.
Format and length
The preparation material can be received as PDF, TXT, or Word files in a list form, summary, or short paragraph relevant to the topics being discussed. It is preferable to send a list of one hundred or so words in total as longer pieces might not be processed efficiently by a corrector during a live meeting.
Notice time
While Verbit’s model can adapt “on the fly”, it is recommended to upload terms to the glossary at least three hours ahead of the meeting and ideally, when the order is being made. If this is not possible and in the case of a live event, it is still better to upload material during the meeting as human correctors adjust the ASR and hence generate a mini-language model supporting the next iterations of words and sentences with a slight delay.
Verbit strongly recommends uploading preparation material with each new order being made to support quality and accuracy.
How to upload files?
Reference to the previous article