prep_for_training.py

#!/usr/bin/env python3 

# Tiro.ehf - 2021 David Erik Mollberg (tiro@tiro.is)
# This is an example data preperation script for the Icelandic lecture corpus build for Kaldi 

import re
import os
from collections import defaultdict


OUT_DIR = '.'

PHONEMES = ['[a]','[ai]','[au]','[c]','[ç]','[cʰ]','[ð]','[ei]','[ɛ]','[f]','[ɣ]','[i]',\
            '[ɪi]','[j]','[n]','[n̥]','[ɲ]','[ŋ]','[œi]','[oi]','[ou]','[ɔ]','[pʰ]','[r]',\
            '[r̥]','[s]','[u]','[ui]','[uɪ]','[v]','[x]','[ʏi]','[θ]','[p]','[l]', '[m̥]', \
            '[l̥]', '[œ]', '[œy]', '[ʏ]', '[g]', '[z]', '[tʰ]', '[ɪ]','[h]', '[ə]', '[y]',\
            '[e]', '[kʰ]', '[k]', '[m]', '[ʀ]','[æ]', '[a:]', '[i:]', '[u:]', '[t]']

def open_lecture_file(file_path:str) -> list:
    '''
    Returns the content of the LECTUREs.tsv file as a list of lists
    with the first line omitted.
    '''
    assert os.path.exists(file_path), 'expected the file LECTURES.tsv to exsist'
    return [line.rstrip().split('\t') for line in open(file_path)][1:]



def init_folders(dir:str) -> None:
    '''
    Creates the output folders
    '''

    for s in ['train', 'dev','eval']:
        if not os.path.exists(os.path.join(dir, s)):
            os.makedirs(os.path.join(dir, s), exist_ok=True)


def convert_milli_to_secs(ms:str) -> str:
    '''
    Convers milliseonds to seconds
    e.g.
    47683 -> 47.68
    '''
    return str(round(float(ms)/1000.0, 2))


def remove_puncuation(s:str, c) -> str:
    '''
    An example of how to prepare the text for use in ASR training
    '''

    for p in PHONEMES:
        if p in s:
            s = s.replace(p, '<unk>')

    s = s.lower()
    s = re.sub('\[hik:{0,1}\s{0,1}(.+?)\]', r"\1", s)
    s = re.sub('\[unk:\s(.+?)\]', r"\1", s)
    s = re.sub('[-/]', ' ', s)
    s = re.sub('[,„“;":\.\'\?\!]','', s)
    s = re.sub('\[unk\]', '<unk>', s)
    s = re.sub('[\[\]]', '', s)
    s = re.sub('µ', 'mu', s)
    s = re.sub('ü', 'ú', s)

    return s


def create_spk2utt(out_dir, utt2spk:str) -> None:
    '''
    Reads the utt2spk file and creates spk2utt
    '''
    content = {}
    for line in utt2spk:
        line = line.split(' ')
        if line[1] not in content:
            content[line[1]] = [line[0]]
        else:
            content[line[1]].append(line[0])


    with open(out_dir, 'w') as f_out:
        for spkd_id, segments in sorted(content.items()):
            f_out.write(spkd_id)
            for segment in sorted(segments):
                f_out.write(' ' + segment)
            f_out.write('\n')


def write_to_file(filename, the_list):
    '''
    Takes in a file name and list of strings and writes out
    '''
    with open(filename, 'w') as f_out:
        for line in the_list:
            f_out.write(line+'\n')


def prep_data() -> None:
    info = open_lecture_file('LECTURES.tsv')    
    data = defaultdict(lambda: defaultdict(list))
    for line in info:
        text_file_path = os.path.join(line[1], line[0]+'.txt')
        assert text_file_path, f'expected {text_file_path} to exsist'

        file_content = [line.rstrip().split('\t') for line in open(text_file_path, encoding='UTF-8')][1:]

        recording_id = f"{line[0]}"
        for segment in file_content:
            
            segment_id = f"{line[1]}_{segment[0]}"
            '''
            Create utt2spk

            segment id - speaker id
            e.g.
            c7a5ce8b-a2ff-4c08-8668-0687fc0f6e97_00013 00003
            c7a5ce8b-a2ff-4c08-8668-0687fc0f6e97_00014 00003
                                ...
            '''         
            data[segment[3]]['utt2spk'].append(f"{segment_id} {line[1]}")
            
            ''' 
            Create text

            segment id - text
            e.g.
            c7a5ce8b-a2ff-4c08-8668-0687fc0f6e97_00013 góðan dag
            c7a5ce8b-a2ff-4c08-8668-0687fc0f6e97_00014 Í þessum fyrirlestri er fjallað um lágmarkspör,
                                ...
            '''
            data[segment[3]]['text'].append(f"{segment_id} {remove_puncuation(segment[4], segment)}")

            '''
            Create segments

            segment id - recording id - start time [sek] - end time [sek]
            e.g.
            c7a5ce8b-a2ff-4c08-8668-0687fc0f6e97_00013 c7a5ce8b-a2ff-4c08-8668-0687fc0f6e97 20.99 23.34
            c7a5ce8b-a2ff-4c08-8668-0687fc0f6e97_00014 c7a5ce8b-a2ff-4c08-8668-0687fc0f6e97 24.03 36.88
                                ...
            '''
            data[segment[3]]['segments'].append(f"{segment_id} {recording_id} {convert_milli_to_secs(segment[1])} {convert_milli_to_secs(segment[2])}")

            '''
            Create wavscp

            segment_id - sox coversion - full path to audio file
            9daf46dc-3ae1-4faa-8546-dbca49f48a22_00070 sox - -c1 -esigned -r 16000 -twav - < ~/kennsluromur-ltp-lectures-new/dev/9daf46dc-3ae1-4faa-8546-dbca49f48a22.wav |
            9daf46dc-3ae1-4faa-8546-dbca49f48a22_00076 sox - -c1 -esigned -r 16000 -twav - < ~/kennsluromur-ltp-lectures-new/dev/9daf46dc-3ae1-4faa-8546-dbca49f48a22.wav |
            '''
            data[segment[3]]['wav.scp'].append(f"{recording_id} sox - -c1 -esigned -r 16000 -twav - < {os.path.join(os.getcwd(), line[1], recording_id+'.wav')} |")

    init_folders(OUT_DIR)

    for the_set in data.keys():
        data[the_set]['wav.scp'] = list(set(data[the_set]['wav.scp']))
        for file in data[the_set].keys():
            sorted_list = sorted(data[the_set][file])
    
            data[the_set][file] = sorted_list
            write_to_file(os.path.join(OUT_DIR, the_set, file), sorted_list)

        create_spk2utt(os.path.join(OUT_DIR, the_set, 'spk2utt'), data[the_set]['utt2spk'])

     


if __name__ == '__main__':

    prep_data()