arfticle-three's Introduction

arfticlethree

Follow my cmusphinx-alignment-example instructions to get cmusphinx set up with my driver class and align.sh.
Get an oral argument audio file (MP3) and transcript text from Oyez.org (e.g. here. From the case page, click on the Argument audio link. Download the MP3, and then click Full Transcript Text and copy and paste the transcript into a similarly named text file. You should have files like 13-553_20141209-argument.mp3 and 13-553_20141209-argument.txt.
Convert the MP3 to a 16khz 16bit mono WAV:

sox 13-553_20141209-argument.mp3 -b 16 13-553_20141209-argument.wav channels 1 rate 16k
Run the transcript parser to segment the transcript by speaker:

python3 parse_transcript.py 13-553_20141209-argument.txt
It won't necessarily get the speakers right. Look at 13-553_20141209-argument.speakers.txt and put #-marks before any line that is not actually a speaker. If you made any change, re-run the script above (it'll look for this file this time and use the list of speakers in the file). The script will write out 13-553_20141209-argument.words.txt which contains just the spoken words.
Run the aligner. This part will take about 2-3 times as long as real time. If the oral argument was an hour, expect the alignment process to take 2-3 hours to finish.

path/to/align.sh 13-553_20141209-argument.wav 13-553_20141209-argument.words.txt > 13-553_20141209-argument.csv
Extract the speaker turn times by running the last script:

python3 get_alignment.py 13-553_20141209-argument.txt > 13-553_20141209-argument.turns.csv

You'll get 13-553_20141209-argument.turns.csv, which is a CSV file where the first column is the name of a speaker and the second and third columns are the start and end times (in miliseconds) of when that speaker was speaking.

Recommend Projects