Creating Custome Scorer DeepSpeech v0.7.1

DeepSpeech
#1

I am running DeepSpeech v0.7.1 on a Raspberry PI4
I am trying to generate a custom scorer since I am using a very small vocabulary set. Research has led me to the following documented under data/lm and generate_package.py.

Prior to running it with all the arguments I ran it without any arguments. This gave me the following:
./generate_package.py
Traceback (most recent call last):
File “./generate_package.py”, line 8, in
import ds_ctcdecoder
ImportError: No module named ds_ctcdecoder

Further investigating led me to try this:
pi@raspberrypi:~ $ pip3 install $(python3 util/taskcluster.py --decoder)
python3: can’t open file ‘util/taskcluster.py’: [Errno 2] No such file or directory

Continuing I tried
pi@raspberrypi:~/DeepSpeech $ pip3 install --upgrade --force-reinstall -e .
Looking in indexes: https://pypi.org/simple, https://www.piwheels.org/simple
Obtaining file:///home/pi/DeepSpeech
ERROR: Could not find a version that satisfies the requirement tensorflow==1.15.2 (from deepspeech-training==0.7.1) (from versions: 0.11.0, 1.12.0, 1.13.1, 1.14.0)
ERROR: No matching distribution found for tensorflow==1.15.2 (from deepspeech-training==0.7.1)

So I guess my question is what are the proper steps for creating a custom scorer?

#2

The problem here is that we don’t (yet?) have an ARM wheel for ds_ctcdecoder.
Do you require producing the LM locally? If not you can generate on another supported device for the time being.

#3

Bummer. I’m fairly new at this. I can attempt to run it on Windows 10 using WSL.
Would this work? Could I use the generated package on my Pi4?

Hank Poletto
hankpoletto62@gmail.com

#4

No need for WSL: https://github.com/mozilla/DeepSpeech/releases/tag/v0.7.1 there are ds_ctcdecoder windows packages

Yes.