Accurate speech-to-text transcriptions are a critical element of Moment exercises + Conversation Simulations. Yet we know accents and dialects can negatively impact transcription accuracy. Custom language models can help!
Your Bright instance automatically comes with a dedicated, company-specific language model that - if you use it - will begin to learn the context of your company's products, transaction types, calls flows and more. The more you give Bright feedback and context, the more accurate the model becomes.
Before we explain how to do this, it's useful to share a simple overview of how language models for transcription actually work. To illustrate we need to (eek!) talk about mathematics.
When your iPhone transcribes a voice note, what it's actually doing is running a probability calculation. It hears what you say, and then picks the most probable outcome.
For example, if you say the sentence "It depends how much money you may owe." Here 'owe' references a financial obligation. A common transcription error could be "It depends how much money you mayo." As in... the condiment!? Why? Because the model (incorrectly) thinks the term 'mayo' is more probable.
How do you fix this? You give the model more context about your company, your preferences, etc. This shifts the probability calculation to be more accurate.
Bright custom language models do the same thing! We allow you to 1) correct transcriptions via the Coach Inbox and/or 2) provide transcripts from your company to build up the platform's knowledge of you and your learners. Here's how:
1) First you'll note the 'Correct Transcription' option for unrated audio recordings that are "In Training" in the Coach Inbox
Note: you cannot edit transcriptions for moments that have been previously rated, or that do not have AI in Training mode.
2) If you click on it you'll see the original transcription + an editable version you can correct
3) Once you make the correction you'll see a confirmation and you'll notice that the learner's color highlights have been removed. This is because the position of the matching learner statements has changed so the highlights may no longer be correct.
4) Now you can refresh the AI rule to run the new transcript against the AI model. This will provide feedback to the learner if the rule is live, otherwise the suggested AI rating will appear as normal in the coach inbox.
It's important to note that Custom Language Models are awesome, but they take time to evolve. Updating 1-2 transcriptions will NOT guarantee that the next transcription will be accurate. As a rough guideline, it will take dozens of updates to see a meaningful difference in transcription.
For faster + better results, speak to your Account Manager about uploading relevant company call transcripts in bulk.