SPRaak Software tool

This software is used to process the given speech data in such way that it resembles telephone speech.

Corpus File

See also this README for further information (section .stm files)

Processing a collection of files in SPRaak requires a corpus (or segmentation) file

A corpus files list the files (or parts thereof) and their content (orthographic description). Files in SPRaak are suposed to have a header (you can specifiy an external header if the file just contains raw data). The header start with the "magic" string ".spr", then continues with a set of "key" "value" value pairs, and ends with a line containing a single "#. The data start directly after the "#" (and newline character).

A corpus file contain ASCII based data (filename + orthographic description + time info + optional extra info), and had the following mandatory fields:

• field1: file name without source path of the databased and without extension; this allows reusing the same corpus file for listing the wav-files or various versions of processed data use - if the next line describes the same file (e.g. when describing each sentence in a seperate line)
• field2: orthograpic description (words are sperated by '_')
• field3+4: begin and end time of the sentence (just a chunk of the signal) time can be
  • start frame (by default 10msec frames) and length in number of frames (integers); this is the default;

  note: the key FSHIFT can be used to redefine the default frame period of 0.01 seconds (10msec)

  • begin time and end time in seconds (floating point value); add the "TIMEBASE CONTINUOUS" to the header if you want to use this convention

  note: use -1 to indicate "till the end of the file" (both when using the frame indices or when using the time in seconds)

• field5... optional fields; you may want to split the data per speaker and add the speaker ID as 5th field (will be needed when doing speaker normalisation)!

Installation

To ease the installation a Dockerfile was made to create a container withholding the SPRaak tool. Just follow the succeeding commands and you will be fine.

Commands

First you need to build an image from that dockerfile with:

docker image build -t $USER/spraak .

Run the container in an interactive mode (terminal) with a volume mounted (most likely the directory in which the speech files are stored).

docker run -ti -v "$PWD/os_volume_path":/container_path $USER/spraak bash

To be able to interact with the spr_ API, run this command:

source /etc/profile

If you want to do a quick check if everything works, just run:

printf '@1+1\nquit\n' | spr_cwr_main
# should print "const ans = 2"

Configuration

Now, configure configure the convert_tel_data.sh script, this is located in the /SPRaak foloder. Within this script the variables dbase, wav_src, and wav_tel_dst must be set to their corresponding values (the nano texteditor is installed).

An example

I set the variables to:

dbase=tel_db
wav_src=/ugent_data/ugent
wav_tel_dst=/ugent_data/ugent_conv_tel

Then I executed the convert_tel_data.sh which led to the following output:

root@77745bdfe6fc:/SPRaak/v1.2.395# ./convert_tel_data.sh 
000000.000868 INFO main_thread.spr_stream_file_open().0 : making directory "/ugent_data/ugent_conv_tel/evaluation/wav/BN"
000127.634697 INFO main_thread.spr_stream_file_open().0 : making directory "/ugent_data/ugent_conv_tel/evaluation/wav/Documentaries"
000329.227871 INFO main_thread.spr_stream_file_open().0 : making directory "/ugent_data/ugent_conv_tel/evaluation/wav/English"
000367.714922 INFO main_thread.spr_stream_file_open().0 : making directory "/ugent_data/ugent_conv_tel/evaluation/wav/Flemish"
000401.465159 INFO main_thread.spr_stream_file_open().0 : making directory "/ugent_data/ugent_conv_tel/evaluation/wav/French"
etc ...

The corpus database file was named "tel_db.cor".