Hi,
Am trying to figure out how to use deep speech in windows/vsstudio but i am having troubles doing the steps, my question is even if i generate libdeepspeech.so using my system, would other people using the application in other systems be able to run?
Your question is not clear. But prebuilt binaries for Windows will arrive soon.
Sorry, if I build the binaries using my system, would I be able to distribute the application using the same binaries or does the binaries need to be build by every different computer
Well, that depends on what you do, I can’t tell what you are going to do, where, and how much redistribuable it can be.
Its going to be for English, american accent, each user would have its own application running locally, for about 40-50 computers
No, if you compile with avx or avx2 just make sure all the others computers supports the instructions you enabled.
What is not working?
Yes we would be using the avx/avx2 since most computers don’t have cuda gpu
Bezel installation was a big hassle, managed to install very easily using chocolatey (for future readers)
problem now is running this command
bazel build -c opt --copt=/arch:AVX --copt=/arch:AVX2 //native_client:libdeepspeech.so
it appears nothing happens, no cpu/gpu/disk/ram usage
edit: got this error
X:\DeepSpeech\tensorflow>bazel build -c opt --copt=/arch:AVX --copt=/arch:AVX2 //native_client:libdeepspeech.so
WARNING: The following rc files are no longer being read, please transfer their contents or import their path into one of the standard rc files:
x:\deepspeech\tensorflow/tools/bazel.rc
Starting local Bazel server and connecting to it…
INFO: Options provided by the client:
Inherited ‘common’ options: --isatty=1 --terminal_columns=161
INFO: Options provided by the client:
‘build’ options: --python_path=C:/Python27/python.exe
INFO: Reading rc options for ‘build’ from x:\deepspeech\tensorflow.tf_configure.bazelrc:
‘build’ options: --action_env PYTHON_BIN_PATH=C:/Python27/python.exe --action_env PYTHON_LIB_PATH=C:/Python27/lib/site-packages --python_path=C:/Python27/python.exe --config=xla --action_env TF_NEED_OPENCL_SYCL=0 --action_env TF_NEED_ROCM=1 --action_env TF_NEED_CUDA=0 --action_env TF_DOWNLOAD_CLANG=0 --config monolithic --copt=-w --host_copt=-w --verbose_failures --distinct_host_configuration=false --experimental_shortened_obj_file_path=true --define=override_eigen_strong_inline=true
INFO: Found applicable config definition build:xla in file x:\deepspeech\tensorflow.tf_configure.bazelrc: --define with_xla_support=true
ERROR: Config value monolithic is not defined in any .rc file
Did you run configure.py?
Yes, tried different options as well trying to figure out atm
updated to python 3.8 with same result
Bazel version? With new version of bazel might not work
bazel version 0.23.2, going to try and downgrade
Try the recommended versions of the readme
It seems to be working with the version [v0.17.2]!
Problem is now C:/users/exterminator/_bazel_exterminator/ky6dhgfz/external/jpeg/BUILD.bazel:401:1: Executing genrule @jpeg//:simd_win_x86_64_assemble failed (Illegal instruction): bash.exe failed: error executing command
then a big chunck of directories appear
then
/usr/bin/bash: line 1: 2276 Illegal instruction bazel-out/x64_windows-opt/bin/external/nasm/nasm.exe -fwin64 -DWIN64 -D__x86_64__ -I $(dirname external/jpeg/simd/x86_64/jccolext-sse2.asm)/ -I $(dirname external/jpeg/simd/nasm/jdct.inc)/ -I $(dirname external/jpeg/simd/nasm/jdct.inc)/…/…/win/ -o $out $(dirname external/jpeg/simd/x86_64/jccolext-sse2.asm)/$(basename ${out%.obj}.asm)
Target //native_client:libdeepspeech.so failed to build
INFO: Elapsed time: 93.769s, Critical Path: 10.69s
INFO: 97 processes: 97 local.
FAILED: Build did NOT complete successfully
You are carefully following the readme? this error is new to me, no idea
Try without avx2? Are you sure that you cpu supports avx2?
Thank you, that was exactly it, after quite a while it worked!
Fixed some config issues on visual studio projects and its now doing recognizing, although, it takes up 2gb ram and it fails to recognize a lot of words
I tested by recording audio from a input device from a show The Office, language english american accent, it fails about 20%/40% of the words easily
You mean from the speaker using the micro or from the windows output? WPF or console? test with librivox recordings. The Windows solutions proven to score WER 8.87%
Using WPF tried from windows output, still trying to fix sound from microphone
If i play the arctic_a0024.wav it did it perfectly, but it was a perfect recording which is not possible in my application which will be recording from a microphone
Also, it takes quite a while to transcribe from real-live audio is this normal behavior?
Remember that the model is not good at handling noise yet, maybe the audio contains laughs or claps?
Yes if you disabled avx2
Yes it does have laughs claps, random background noise which really hurts the recognition, on a clean audio its pretty good
Unfortunately my cpu doesn’t have avx2
Would the dataset from mozilla common voice be better at handling background noise?