2021-08-25 kaldi

Kaldi Learning Record

Two parts in speech recognition

1. Prenounciation model
2. Grammar model

Chapter 2: Basic of Kaldi

\egs\yesno

1. perl script for multi jobs and print the output to the log at the same time utils/run.pl JOB=1:5 log command
2. run.sh the main script for handling part1 and part2.

Basic Structure of one model:

1. cmd.sh : Config how the project is handed. If seg is used, queue.pl; if single pc is used, run.pl; if NFS and ssh is configured mkdir .queue && touch machines && for name in all_names export; do echo ${name} > .queue/machines; done; if used self-wrriten .pl, configure as the single pc but replace the .pl script
2. path.sh : Config the kaldi root path and path. If self preferred paths are applied, then specify in this script.
3. RESULTS save the models results
4. local saves all the scripts
5. conf saves all the configure file
6. steps, utils are commonly used scripts

Chapter 3

Data pre-processing

1. Kalda uses tables to store and find objects. There are list tables and save tables. List tables have .sc p they look like name \root\absolute\name or in UNIX use pipeline like this name gzip -c \root\absulote\name.zip | the last | means to pass this to the stdout. If there are program needs this pipeline just add program like this | program then the program would catch the pipeline result. This way can save memory. So the list tables are more like a index.
   Save tables are like .ark. and has contents like This is ark content1 \nThis is ark content 2 This table stores have special usage in binary files. When the tables are stored as ‘b’ the making is ’\0B . In binary tables we could add offsets and apply slicing like :17[:,0:8] menas to get byte from 17 and get the former 9 dimensions of data.
2. Using tables to input and output:
   e.g wav-to-duration this is a kaldi’s own program and is used to generate spk2dur.apk (the time duration for each text)
3. Tages for tables:

Writing features:

1. Table types: `scp` or `ark` if two types of tables are generated at the same time , must be ark scp ordered.
2. Binary mode: `b`  save the apk tables as binary and is default
3. Text mode: