coraal-asr.py

# coding=utf-8
# Copyright 2021 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Common Voice Dataset"""

import json
import os
from copy import deepcopy
import re
import unicodedata
from more_itertools import windowed
import datasets

_CITATION = """\
"""

_DESCRIPTION = """\
coraalを音声認識した誤り訂正用データセット
"""
_HOMEPAGE = ""
_LICENSE = ""

URLS = {
    "v1": {
        "text": "https://huggingface.co/datasets/Padomin/coraal-asr/resolve/main/coraal-asr.tar.gz",
    },
    "v2": {
        "text": "https://huggingface.co/datasets/Padomin/coraal-asr/resolve/main/coraal-asr-v2.tar.gz",
    },
    "ctc-large": {
        "text": "https://huggingface.co/datasets/Padomin/coraal-asr/resolve/main/coraal-ctc-large.tar.gz",
    },
    "xlsr": {
        "text": "https://huggingface.co/datasets/Padomin/coraal-asr/resolve/main/coraal-xlsr.tar.gz",
    },
    "whisper-small": {
        "text": "https://huggingface.co/datasets/Padomin/coraal-asr/resolve/main/coraal-whisper-small.tar.gz",
    }
}


class coraal_asr_config(datasets.BuilderConfig):
    def __init__(self, n_fronts=0, n_bodies=1, n_rears=0, front_prefix='front:\n', body_prefix='body:\n', rear_prefix='rear:\n', **kwargs):
        super(coraal_asr_config, self).__init__(**kwargs)
        self.n_fronts = n_fronts
        self.n_bodies = n_bodies
        self.n_rears = n_rears
        self.front_prefix = front_prefix
        self.body_prefix = body_prefix
        self.rear_prefix = rear_prefix

class coraal_asr(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("0.2.0")
    BUILDER_CONFIGS = [
        coraal_asr_config(name="v1", version=VERSION),
        coraal_asr_config(name="v2", version=VERSION),
        coraal_asr_config(name="ctc-large", version=VERSION),
        coraal_asr_config(name="xlsr", version=VERSION),
        coraal_asr_config(name="whisper-small", version=VERSION),
        coraal_asr_config(name="ctc-large-oracle", version=VERSION),
    ]
    DEFAULT_CONFIG_NAME = "ctc-large"  # It's not mandatory to have a default configuration. Just use one if it make sense.
    BUILDER_CONFIG_CLASS = coraal_asr_config

    def _info(self):
        feature_dict = {
                "text": datasets.Value("string"),
                "text_asr": datasets.Value("string"),
                "src": datasets.Value("string"),
                "tgt": datasets.Value("string"),
                "id": datasets.Value("string")
        }

        features = datasets.Features(feature_dict)
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            supervised_keys=None,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        """Returns SplitGenerators."""
        if "v1" in self.config.name:
            urls = deepcopy(URLS["v1"])
        if "v2" in self.config.name:
            urls = deepcopy(URLS["v2"])
        if "ctc-large" in self.config.name:
            urls = deepcopy(URLS["ctc-large"])
        if "xlsr" in self.config.name:
            urls = deepcopy(URLS["xlsr"])
        if "whisper-small" in self.config.name:
            urls = deepcopy(URLS["whisper-small"])
        if "ctc-large-oracle" in self.config.name:
            urls = deepcopy(URLS["ctc-large"])
            
        dl_path = dl_manager.download_and_extract(urls)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={
                    "filepath": os.path.join(dl_path["text"], "train.jsonl"),
                    "split": "train",
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.TEST,
                gen_kwargs={
                    "filepath": os.path.join(dl_path["text"], "test.jsonl"),
                    "split": "test",
                },
            ),
            datasets.SplitGenerator(
                name=datasets.Split.VALIDATION,
                gen_kwargs={
                    "filepath": os.path.join(dl_path["text"], "validation.jsonl"),
                    "split": "validation",
                },
            ),
        ]

    def _generate_examples(self, filepath, split):
        """Yields examples."""
        id_ = 0
        with open(filepath, encoding="utf-8") as f:
            for line in f:
                doc = json.loads(line)
                utterances = doc['utterances']
                # divide text and asr
                texts_asr = [utt['asr'] for utt in utterances]
                texts = [utt['text'] for utt in utterances]
                # window considering front and rear contexts
                if split == "train":
                    windowed_texts_asr = windowed([''] * self.config.n_fronts + texts_asr + [''] * self.config.n_rears, self.config.n_bodies + self.config.n_fronts + self.config.n_rears)
                    windowed_oracles = windowed([''] * self.config.n_fronts + texts + [''] * self.config.n_rears, self.config.n_bodies + self.config.n_fronts + self.config.n_rears)
                    windowed_texts = windowed(texts, self.config.n_bodies)
                else:
                    windowed_texts_asr = windowed([''] * self.config.n_fronts + texts_asr + [''] * self.config.n_rears, self.config.n_bodies + self.config.n_fronts + self.config.n_rears, fillvalue='', step=self.config.n_bodies)
                    windowed_oracles = windowed([''] * self.config.n_fronts + texts + [''] * self.config.n_rears, self.config.n_bodies + self.config.n_fronts + self.config.n_rears, fillvalue='', step=self.config.n_bodies)
                    windowed_texts = windowed(texts, self.config.n_bodies, fillvalue='', step=self.config.n_bodies)
                
                for text_asr, text, oracle, utt in zip(windowed_texts_asr, windowed_texts, windowed_oracles, utterances):
                    src = ''
                    if self.config.n_fronts > 0:
                        src += self.config.front_prefix
                        if "oracle" in self.config.name:
                            src += '\n'.join(oracle[:self.config.n_fronts])
                        else:
                            src += '\n'.join(text_asr[:self.config.n_fronts])
                        src += '\n'
                    src += self.config.body_prefix
                    src += '\n'.join(text_asr[self.config.n_fronts:self.config.n_fronts + self.config.n_bodies])
                    if self.config.n_rears > 0:
                        src += '\n' + self.config.rear_prefix
                        if "oracle" in self.config.name:
                            src += '\n'.join(oracle[self.config.n_fronts + self.config.n_bodies:])
                        else:
                            src += '\n'.join(text_asr[self.config.n_fronts + self.config.n_bodies:])
                    tgt = '\n'.join(text)
                    
                    data = {
                        "text": utt["text"],
                        "text_asr": utt["asr"],
                        'src': src,
                        'tgt': tgt,
                        'id': doc["id"],
                    }
                    
                    yield id_, data
                    
                    id_ += 1