KaldiDecoder (Source Compilation)
These instructions are for KaldiDecoder version 2.5.0 or later.
Compilation Environment
- OS
- Ubuntu 16.04 Xenial (64bit)
- Ubuntu 18.04 Bionic (64bit)
- Ubuntu 20.04 Focal (64bit)
- Ubuntu 22.04 Jammy (64bit)
Notes: KaldiDecoder uses the Kaldi libraries, so it can only be built in an environment supported by the Kaldi officially site (http://kaldi-asr.org/).
Installation instruction
Step.1 Get build tools and dependent libraries.
sudo apt update sudo apt install git automake autoconf libtool cmake cmake-extras build-essential libopenblas-base libopenblas-dev gfortran liblapack-dev liblapacke-dev zlib1g-dev sox subversion
For Ubuntu 20.04 or later users, you will need to additionally install the python2.7 package.
sudo apt install python2.7
Step.2 Get the official Kaldi source code.
cd ~/ mkdir YOUR_DIR cd YOUR_DIR git clone https://github.com/kaldi-asr/kaldi.git cd kaldi git checkout COMMIT_ID
The binary package provided by us was built based on the commit ID below. If you change the base Kaldi version, the build may fail.
- kaldidecoder-hark: 2.5.0-openblas, 3.0.0-openblas, 3.0.5-openblas
- COMMIT_ID = 4571f47f84 (Kaldi 5.4.173)
- kaldidecoder-hark: 3.1.0-openblas, 3.1.2.1-openblas, (and internal release 3.1.3-openblas, 3.1.4-openblas, 3.1.5-openblas, 3.1.6-openblas, 3.1.7-openblas, 3.2.0-openblas)
- COMMIT_ID = c0a555e4ce (Kaldi 5.5.214)
- kaldidecoder-hark: 3.4.0-openblas
- COMMIT_ID = 7a50987e7 (Kaldi 5.5.636 : Although it is possible to build it, it was not adopted in the packaged version.)
- COMMIT_ID = 1a233a11db (The newer commit we saw was committed on 2022/05/26 01:03:14 JST)
Step.3 Build Kaldi’s external tool (and dependent library).
cd tools extras/check_dependencies.sh #### The following message is displayed but there is no problem. Because we use openblas. # You should probably do: # sudo apt-get install libatlas3-base make
Step.4 Build Kaldi with the set up to use openblas library.
For Ubuntu 16.04 users.
cd ../src ./configure --mathlib=OPENBLAS --openblas-root=/usr make clean -j CORES make depend -j CORES make -j CORES
- CORES : How many cores do you have
-
e.g.) 4
Note: About 2GB of memory is required for each core when compiling. Please set the number of cores considering free memory capacity.
For Ubuntu 18.04 or later users, the following workaround is required because the openblas library path was changed.
cd ../src mkdir fakeroot ln -s /usr/include/x86_64-linux-gnu/ fakeroot/include ln -s /usr/lib/x86_64-linux-gnu/ fakeroot/lib ./configure --mathlib=OPENBLAS --openblas-root=fakeroot make clean -j CORES make depend -j CORES make -j CORES
- CORES : How many cores do you have
-
e.g.) 4
Note: About 2GB of memory is required for each core when compiling. Please set the number of cores considering free memory capacity.
Step5. Build KaldiDecoder.
cd ../ wget http://archive.hark.jp/harkrepos/dists/DISTRO/non-free/source/kaldidecoder-hark_HARK_VER.tar.xz tar -Jxvf kaldidecoder-hark_HARK_VER.tar.xz cd kaldidecoder3 mkdir build cd build cmake .. -DOPENBLAS_ROOT_DIR:STRING=/usr -DCMAKE_VERBOSE_MAKEFILE:BOOL=TRUE make sudo make install
If you want to know which HARK_VER that can be downloaded, please access directly from your browser at the following location.
- For Ubuntu 16.04 (xenial) users.
-
http://archive.hark.jp/harkrepos/dists/xenial/non-free/source/
- For Ubuntu 18.04 (bionic) users.
-
http://archive.hark.jp/harkrepos/dists/bionic/non-free/source/
- For Ubuntu 20.04 (focal) users.
-
http://archive.hark.jp/harkrepos/dists/focal/non-free/source/
- For Ubuntu 22.04 (jammy) users.
-
http://archive.hark.jp/harkrepos/dists/jammy/non-free/source/
- DISTRO : Ubuntu distribution
-
e.g.) xenial, bionic, focal, jammy, etc…
Q&A
Q.1 Is there a way to enable CUDA?
A.1 If you are using 3.1.0-openblas or later, the answer is Yes, but do so at your own risk.
Steps 1 to 3 are the same. For installation of CUDA toolkit please refer to nVidia official installation instructions.
- nVidia CUDA Toolkit
-
https://developer.nvidia.com/cuda-downloads
Step.4 Build Kaldi with the set up to use openblas library. (with CUDA)
For Ubuntu 16.04 users.
cd ../src ./configure --mathlib=OPENBLAS --openblas-root=/usr --use-cuda --cudatk-dir=/usr/local/cuda make clean -j CORES make depend -j CORES make -j CORES
Only the arguments of '--use-cuda' and '--cudatk-dir=PATH' have been added by the './configure' command.
For Ubuntu 18.04 or later users, the following workaround is required because the openblas library path was changed.
cd ../src mkdir fakeroot ln -s /usr/include/x86_64-linux-gnu/ fakeroot/include ln -s /usr/lib/x86_64-linux-gnu/ fakeroot/lib ./configure --mathlib=OPENBLAS --openblas-root=fakeroot --use-cuda --cudatk-dir=/usr/local/cuda make clean -j CORES make depend -j CORES make -j CORES
Only the arguments of '--use-cuda' and '--cudatk-dir=PATH' have been added by the './configure' command.
Step5. Build KaldiDecoder. (with CUDA)
cd ../ wget http://archive.hark.jp/harkrepos/dists/DISTRO/non-free/source/kaldidecoder-hark_HARK_VER.tar.xz tar -Jxvf kaldidecoder-hark_HARK_VER.tar.xz cd kaldidecoder3 mkdir build cd build cmake .. -DOPENBLAS_ROOT_DIR:STRING=/usr -DCMAKE_VERBOSE_MAKEFILE:BOOL=TRUE -DUSE_CUDA=ON make sudo make install
Only the arguments of '-DUSE_CUDA=ON' have been added by the 'cmake' command.
If you have multiple versions of CUDA installed and you want to set your PATH:
cmake .. -DOPENBLAS_ROOT_DIR:STRING=/usr -DCMAKE_VERBOSE_MAKEFILE:BOOL=TRUE -DUSE_CUDA=ON -DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-9.0
Make sure that the PATH specified here is the same as the PATH specified in Step 4.
