|TalkBank||AWS and ASR|
We have recently been testing out the use of automatic speech recognition (ASR) through various online web services. At this point, the service that has worked best for us is Amazon Web Services (AWS). We used AWS ASR to recognize the words in 24 20-minutes clinical interviews with control participants for our work on language in dementia. The output from AWS was accurate at the word level, but needed a lot of additional massaging to get into CHAT format. Fortunately, much of this additional work can be partially automated using the methods described in this guide . We estimate that use of the method saves at least 30% of time in transcription, particularly once one gets used to the flow of the steps.
Our tests so far have been with adult control participants without language disorders. How well this will work with children or participants with language disorders is not yet clear.
The initial conversion from AWS format to CHAT relies on this Python script which must be run in Python 3.10.x whose installer can be downloaded from www.python.org
We have also created an asr.cut file for use by the CHSTRING program, as described in the guide.
We are monitoring several other methods, including use of the SpeechKitchen and ESPNet systems, both developed at CMU. We welcome input regarding suggestions for other methods to achieve accurate ASR.