1. Help Center
  2. Skills Coaching + AI
  3. Language Models + Transcription

How to Prepare Transcripts for Upload to Your Custom Language Model

You're ready to level up your learner's AI-Coaching Experience, but how do you upload call transcripts?

Your Bright Account Manager will handle the upload itself for you, but there's plenty that you can do to make sure that the process is smooth and the impact is tangible.

What calls should you collect?

You should aim to collect at least 50-100 calls for each line of business that you are supporting in Bright to achieve a more tailored language model - and more accurate transcription of your learners' recorded responses. For example, if you have 3 departments - sales, customer care and scheduling - in Bright, you would want between 150-300 calls, with equal representation from each. 

The higher the volume of words ingested into the model, the bigger the shift in the NLP's output!

The calls do not need to be a perfect QA score, but having a good representation of strong calls will certainly help as they may use some words or phrases that poor calls are missing.

So you need to ask for transcripts from a vendor. What should you put in your request?

If your call software does not offer bulk export of transcripts, you may choose to work directly with the vendor to pull down a few hundred transcripts. If so, you can use or slightly modify the following sample request:

We would like to pull down a cross-section of call types, about XX total calls, in a .txt file. We only need the transcription and not the audio for the calls. This is for a project with a training vendor of ours. Please include a sample of calls across business lines and groups.

Note: if you have multiple locations or groups in the same line of business, you may wish to ensure that each location is included in your sample as well.

What do you do with the calls once you have the transcriptions?

  1. All files need to be .txt files.
  2. Make sure that the calls do not have any protected information, such as PII, PCI, or PHI.
  3. Review for typos or errors
  4. Ensure that the transcription of the call matches what was said by the agent and customer - we want a high volume of the right words (e.g. if you want the model to know your learners will say "deductible" or "may owe," then these need to appear correctly in the files we load)
  5. Once all the above steps have been completed, share the files with your account manager via your shared file space or an emailed zip file

What do I do if certain words are never transcribed correctly? 

We also have the ability to upload the phonetic spelling of certain words. These tend to be proper nouns (like your company or customer names) that are regularly transcribed to different common words. To learn how to add this to your custom language model, see this help article