This repository contains a recipe for training an automatic speech recognition (ASR) system using the TIDIGITS database. The recipe is entirely Julia-flavoured and uses following packages (among others):
- Flux as ML library
- FiniteStateTransducers for WFST compositions
- HMMGradients for maximum likelihood training
Currently the training runs only on CPU and employs a simple greedy decoder.
Run julia --project -e 'using Pkg; Pkg.instantiate()' to install all the dependencies.
For the live demo install sox.
Open a Julia terminal with julia --project and type include("demo.jl") to try out the ASR with your own voice. A model trained with configuration 2b (see below) is already present in this repository.
Specify your current configuration in the folder conf.
The configuration files are loaded from the folder conf/mysetup/.
This folder must contain the following files:
feat_conf.jlfor feature extractionmodel_conf.jlfor model and optimisation parameters (hyperparameters) A couple of setups are present in this repository for reference in the folderconf. Currently a TDNN/ConvNet is used as acoustic model.
Set in your shell environment the path TIDIGITS_PATH=\your\path\to\tidigits.
If you're using SGE set the command flags in CPU_CMD, i.e. the queue options.
This can be done e.g. by running source env.sh before lunching Julia, where env.sh is a script that export these variables.
Alternatively, the environment variables can be specified directly in the REPL.
Run julia --project prepare_data.jl --conf 2a to extract feature and prepare training data using the configuration 2a.
Features and transcriptions will be saved in the folder data/uuid/.
Here uuid is linked to feat_conf.jl file, meaning that if you create a new model_conf.jl without modifying feature extraction you don't need to run data preparation twice.
If SGE grid is available add the flag --nj N to split the work into N jobs.
For the moment HMM configuration is fixed in wfsts.jl with a phone based 2-state HMM.
Training is performed running the script julia --project train.jl --conf 2a.
Notice that if you're just experimenting it is more convenient to run the experiment from Julia's REPL.
$ julia --project
julia> include("train.jl")
Modify the conf by changing the default in the ArgParse table.
Run the script eval.jl to calculate Word Error Rates (WER) and Phone Error Rate (PER).
This work was developed under the supsrvision of Prof. Dr. Hervé Bourlard and supported by the Swiss National Science Foundation under the project "Sparse and hierarchical Structures for Speech Modeling" (SHISSM) (no. 200021.175589).