30
Process of Language Model Improvements
DeploymentModel TuningModel TrainingTranscriptionSecure HandoverCall Download
Call Download / Selection
▪ 100 hours of handset audio in the relevant
language
▪ Statistical dialect representation of the trading
floor and form minimum 50 traders
▪ Best practices to select audio data from a 2
week window
▪ Files to be downloaded in a WAV format
Secure Handover
▪ NICE & Customer signs a NDA and Data
Protection Agreement
▪ Customer hands over data to NICE though
agreed secure channel
▪ Data stored on secured NICE servers (only
registered persons have access)
▪ Data securely deleted from NICE servers after
usage (under Customers supervision)
Transcription (Manual)
▪ All calls that are handed over needs to be
manually transcribed and tagged with events
(see transcription method documentation)
▪ Transcription of these calls happens under
strict supervision of NICE by transcription
provider (option to replace by customer
certified provider)
▪ Transcription is checked to achieve 99%
quality.
Model Training
▪ After transcription is finished then the NICE
Language Model team take the audio data and
transcription data (leaving 10% for control
testing)
▪ The data is feed into a learning machine that
scans the two sources and creates an acoustic
model (if needed) and a language model
(including a dictionary and NNLP packs)
Model Tuning
▪ Based on the results of the testing the model
can be further tuned, if the expected results
have not been reached
▪ Acoustic and language models can be tuned
manually or additional data sets can be used
Deployment
▪ When the required quality has been reached
the models are ready for deployment