@CIOL_Linguists
FEATURES
Linguists also have to check that the
training data is consistent, for example by
developing and enforcing guidelines on how
colloquial speech should be transcribed. For
instance they could determine that it is more
important for a transcribed sentence to follow
formal standard grammar than to
accurately depict the speech recording.
In this case, the transcription might
change a sentence such as 'they was gonna
go out' to formal English 'they were going to
go out'. The solution should balance the ease
of processing for the machine learning
algorithm with the importance of preserving
the speech in its original form.
This is a growing area of work and the
Conversation Design Institute's online 'AI
Trainer' course and exam now provide
an industry-recognised certificate in training
language models for anyone interested
in diversifying.
Determining emotion
Language specialists also work on other
features of ASR systems, such as establishing
the emotions of the speaker, known as
'sentiment labelling'. Though ascribing an
emotion to speech seems intuitive, the
challenge is to consider the culture and
background of the speaker, as well as the
biases of the people transcribing.
Professionals with a background in linguistics
or sociolinguistics can use their expertise to
notice specific linguistic features in the
speech recording that lead towards an
accurate sentiment label.
An expert could identify that the speaker
is using their second language and
determine what their native language is.
sts is crucial
©
SHUTTERSTOCK
