|
""" Babelbox Voice Dataset""" |
|
|
|
import os |
|
import csv |
|
import codecs |
|
import datasets |
|
from typing import List |
|
from pathlib import Path |
|
from tqdm import tqdm |
|
import torchaudio |
|
from io import BytesIO |
|
import pydub |
|
from pydub import AudioSegment |
|
from pydub.silence import detect_leading_silence |
|
import numpy as np |
|
logger = datasets.logging.get_logger(__name__) |
|
|
|
_CITATION = """\ |
|
@inproceedings{babelboxvoice:2022, |
|
author = {Andersson, O. and Bjelkenhed, M. and Bielsa, M. et al}, |
|
title = {Babelbox Voice: A Speech Corpus for training Whisper}, |
|
year = 2022 |
|
} |
|
""" |
|
|
|
_HF_REPO_PATH = "https://huggingface.co/datasets/babelbox/babelbox_voice/" |
|
|
|
class BabelboxVoiceConfig(datasets.BuilderConfig): |
|
"""BuilderConfig for BabelboxVoice.""" |
|
|
|
def __init__(self, name, version, **kwargs): |
|
self.name = name |
|
self.version = version |
|
self.features = kwargs.pop("features", None) |
|
self.description = kwargs.pop("description", None) |
|
self.data_url = kwargs.pop("data_url", None) |
|
self.nb_data_shards = kwargs.pop("nb_data_shards", None) |
|
self.metadata_url = kwargs.pop("metadata_url", None) |
|
|
|
description = ( |
|
f"Babelbox Voice speech to text dataset." |
|
) |
|
super(BabelboxVoiceConfig, self).__init__( |
|
name=name, |
|
version=version, |
|
**kwargs, |
|
) |
|
|
|
|
|
class BabelboxVoice(datasets.GeneratorBasedBuilder): |
|
|
|
VERSION = datasets.Version("1.0.0") |
|
|
|
BUILDER_CONFIGS = [ |
|
BabelboxVoiceConfig( |
|
name="nst", |
|
version=VERSION, |
|
description="This part of Babel Voice includes data from National Library of Norway", |
|
features=["path", "audio", "sentence"], |
|
data_url= _HF_REPO_PATH + "resolve/main/data/nst/nst-data-{:0>3d}.tar.gz", |
|
nb_data_shards = 42, |
|
metadata_url= _HF_REPO_PATH + "resolve/main/data/nst/metadata.tar.gz" |
|
) |
|
] |
|
|
|
DEFAULT_CONFIG_NAME = "nst" |
|
|
|
def _info(self): |
|
description = ( |
|
"Babelbox Voice is an initiative to help teach machines how real people speak. " |
|
) |
|
if self.config.name == "nst": |
|
features = datasets.Features( |
|
{ |
|
"path": datasets.Value("string"), |
|
"audio": datasets.features.Audio(sampling_rate=16_000), |
|
"text": datasets.Value("binary"), |
|
"speaker_id": datasets.Value("string"), |
|
"sex": datasets.Value("string"), |
|
"accent": datasets.Value("string"), |
|
} |
|
) |
|
else: |
|
features = datasets.Features( |
|
{ |
|
"path": datasets.Value("string"), |
|
"audio": datasets.features.Audio(sampling_rate=16_000), |
|
"text": datasets.Value("binary"), |
|
"speaker_id": datasets.Value("string"), |
|
"sex": datasets.Value("string"), |
|
"accent": datasets.Value("string"), |
|
} |
|
) |
|
|
|
return datasets.DatasetInfo( |
|
description=description, |
|
features=features, |
|
supervised_keys=None, |
|
version=self.config.version |
|
) |
|
|
|
def get_metadata(self, dl_manager, metadata_url): |
|
if metadata_url == None: return None |
|
|
|
metadata_path = dl_manager.download(metadata_url) |
|
local_extracted_metadata_path = dl_manager.extract(metadata_path) if not dl_manager.is_streaming else None |
|
|
|
|
|
def clean_sentence(sentence): |
|
return (sentence |
|
.replace("\\Komma", "",) |
|
.replace("\\Punkt", "") |
|
.replace("\\Utropstecken", "") |
|
.replace("\\Frågetecken", "")) |
|
|
|
metadata_archive = dl_manager.iter_archive(metadata_path) |
|
metadata = {} |
|
for path, file in metadata_archive: |
|
reader = csv.DictReader(codecs.iterdecode(file, 'utf-8')) |
|
for row in tqdm(reader, desc="Reading metadata..."): |
|
filename = row['filename_channel_1'] |
|
metadata_item = { |
|
'sentence': clean_sentence(row['text']), |
|
'speaker_id': row['Speaker_ID'], |
|
'sex': row['Sex'], |
|
'accent': row['Region_of_Youth'] |
|
} |
|
|
|
metadata[filename] = metadata_item |
|
|
|
return metadata |
|
|
|
|
|
def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]: |
|
|
|
download_urls = [self.config.data_url.format(i) for i in range(1, self.config.nb_data_shards + 1) ] |
|
|
|
archive_paths = dl_manager.download(download_urls) |
|
local_extracted_archive_paths = dl_manager.extract(archive_paths) if not dl_manager.is_streaming else {} |
|
|
|
metadata = self.get_metadata(dl_manager, self.config.metadata_url) |
|
|
|
return [ |
|
datasets.SplitGenerator(name=datasets.Split.TRAIN, |
|
gen_kwargs={ |
|
"local_extracted_archive_paths": local_extracted_archive_paths, |
|
"archives": [dl_manager.iter_archive(path) for path in archive_paths], |
|
"metadata": metadata |
|
}) |
|
] |
|
|
|
def _generate_examples(self, local_extracted_archive_paths, archives, metadata): |
|
|
|
sampling_rate = 16000 |
|
|
|
def get_audiosegment(array): |
|
byte_array = np.int16(array * 2 ** 15).tobytes() |
|
audio_segment = pydub.AudioSegment(byte_array, frame_rate=16000, sample_width=2, channels=1) |
|
return audio_segment |
|
|
|
def get_leading_and_trailing_silence(audio_array): |
|
audio_segment = get_audiosegment(audio_array) |
|
|
|
leading = detect_leading_silence(audio_segment) |
|
trailing = detect_leading_silence(audio_segment.reverse()) |
|
return leading, trailing |
|
|
|
def get_timestamp(audio_array): |
|
leading_silence, trailing_silence = get_leading_and_trailing_silence(audio_array) |
|
start_len = leading_silence / 1000 |
|
start_time = round(start_len / 2, 2) * 2 |
|
|
|
end_len = (len(audio_array) / sampling_rate) - (trailing_silence / 1000) |
|
end_time = round(end_len / 2, 2) * 2 |
|
return (start_time, end_time) |
|
|
|
|
|
def filter_sentence(sentence): |
|
if "... tyst under denna inspelning ..." in sentence: return False |
|
return True |
|
|
|
def get_audio(file): |
|
file_buf = BytesIO(file.read()) |
|
array, audio_sr = torchaudio.load(file_buf, format="wav") |
|
return array[0], audio_sr |
|
|
|
for i, audio_archive in enumerate(archives): |
|
for path, file in audio_archive: |
|
if local_extracted_archive_paths == False: |
|
path = os.path.join(local_extracted_archive_paths[i], path) |
|
|
|
metadata_item = metadata[path] |
|
|
|
audio_array, audio_sr = get_audio(file) |
|
|
|
if filter_sentence(metadata_item['sentence']) == False: continue |
|
|
|
audio_len = len(audio_array) / sampling_rate |
|
audio_len = round(audio_len / 2, 2) * 2 |
|
|
|
print(get_leading_and_trailing_silence(audio_array)) |
|
|
|
text = { |
|
"text": metadata_item['sentence'], |
|
"offsets": [ |
|
{"text": metadata_item['sentence'], "timestamp": get_timestamp(audio_array) } |
|
] |
|
} |
|
|
|
|
|
result = { |
|
'path' : path, |
|
'audio': {"path": path, "array": audio_array, "sampling_rate": sampling_rate}, |
|
'text' : text, |
|
'speaker_id' : metadata_item['speaker_id'], |
|
'sex' : metadata_item['sex'], |
|
'accent' : metadata_item['accent'] |
|
} |
|
|
|
yield path, result |
|
|
|
|