train

Training acoustic models with CMU Sphinx

According to CMUSphinx's tutorial AM, the directory tree for new projects must follow the structure below:

             my_db_dir/
                 │ 
  .--------------:--------------.
  │                             │                        
 etc/                          wav/                                     
  ├─ my_db.dic                  ├─ spkr_1/                     
  ├─ my_db.phone                │     ├─ s1_file_1.wav                     
  ├─ my_db.lm                   │     ├─ s1_file_2.wav         
  ├─ my_db.filler               │     └─ s1_file_n.wav              
  ├─ my_db_train.fileids        ├─ spkr_2/                
  ├─ my_db_train.transcription  │     ├─ s2_file_1.wav             
  ├─ my_db_test.fileids         │     ├─ s2_file_2.wav                     
  └─ my_db_test.transcription   │     └─ s2_file_n.wav                   
                                └─ spkr_n/           
                                      ├─ sn_file_1.wav 
                                      ├─ sn_file_2.wav 
                                      └─ sn_file_n.wav 

These scripts cover the "Data Preparation" section of CMU Sphinx's official AM training tutorial.

• fb\_00\_create\_envtree.sh: This script creates the directory structure shown above, except the spkr_X inside the wav folder. Notice that the data-dependent files (inside the etc dir), although created, they DO NOT have any content yet. IOW, they're only initialized as empty files. A stupid choice of ours. But this scripts also checks for dependencies that must be installed before running fb_01 and fb_02, such as sox and wget.
• fb\_01\_split\_train\_test.sh: This script fulfills the fileids and transcriptions files in etc/ dir. The data is divided as training set and test set, and the files within the dirs are data-dependent. The folders wav/spkr_X contain symbolic links to the actual wav-transcription base dir.
• fb\_02\_define\_etclang.sh: This script specially fulfills the files inside my_db_dir/etc dir: .dic, .filler, .phone, and .lm. A dependency is our g2p software, which must be previously downloaded/cloned from https://gitlab.com/fb-nlp/nlp-generator.git.

The next steps will then be (please refer to the section "Setting up the training scripts" for details):

• run sphinxtrain -t my_db_dir setup inside your project dir
• edit the recently created etc/sphinx_train.cfg file
• run sphinxtrain run to begin the AM train.

Grupo FalaBrasil (2020) - https://ufpafalabrasil.gitlab.io/
Universidade Federal do Pará (UFPA) - https://portal.ufpa.br/
Cassio Batista - https://cassota.gitlab.io/