Training in progress, step 500

.gitignore

@@ -0,0 +1 @@
+checkpoint-*/

added_tokens.json

@@ -0,0 +1 @@
+{"<s>": 1205, "</s>": 1206}

config.json

@@ -0,0 +1,107 @@
+{
+  "_name_or_path": "facebook/wav2vec2-xls-r-1b",
+  "activation_dropout": 0.1,
+  "adapter_kernel_size": 3,
+  "adapter_stride": 2,
+  "add_adapter": false,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "codevector_dim": 1024,
+  "contrastive_logits_temperature": 0.1,
+  "conv_bias": true,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "mean",
+  "ctc_zero_infinity": false,
+  "diversity_loss_weight": 0.1,
+  "do_stable_layer_norm": true,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_dropout": 0.0,
+  "feat_extract_norm": "layer",
+  "feat_proj_dropout": 0.0,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.0,
+  "hidden_size": 1280,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.0,
+  "mask_feature_length": 64,
+  "mask_feature_min_masks": 0,
+  "mask_feature_prob": 0.25,
+  "mask_time_length": 10,
+  "mask_time_min_masks": 2,
+  "mask_time_prob": 0.75,
+  "model_type": "wav2vec2",
+  "num_adapter_layers": 3,
+  "num_attention_heads": 16,
+  "num_codevector_groups": 2,
+  "num_codevectors_per_group": 320,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 48,
+  "num_negatives": 100,
+  "output_hidden_size": 1280,
+  "pad_token_id": 1204,
+  "proj_codevector_dim": 1024,
+  "tdnn_dilation": [
+    1,
+    2,
+    3,
+    1,
+    1
+  ],
+  "tdnn_dim": [
+    512,
+    512,
+    512,
+    512,
+    1500
+  ],
+  "tdnn_kernel": [
+    5,
+    3,
+    3,
+    1,
+    1
+  ],
+  "torch_dtype": "float32",
+  "transformers_version": "4.17.0.dev0",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 1207,
+  "xvector_output_dim": 512
+}

preprocessor_config.json

@@ -0,0 +1,9 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0,
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:450eabd0a5373e1ecd8ce2213749283b3b63609903c85ed7ae272817327b3b59
+size 3856497393

run.sh

@@ -0,0 +1,34 @@
+python run_speech_recognition_ctc.py \
+	--dataset_name="kresnik/zeroth_korean" \
+	--model_name_or_path="facebook/wav2vec2-xls-r-1b" \
+	--dataset_config_name="clean" \
+	--output_dir="./" \
+	--overwrite_output_dir \
+	--num_train_epochs="50" \
+	--per_device_train_batch_size="8" \
+	--per_device_eval_batch_size="8" \
+	--gradient_accumulation_steps="4" \
+	--3="7.5e-5" \
+	--warmup_steps="2000" \
+	--length_column_name="input_length" \
+	--evaluation_strategy="steps" \
+	--text_column_name="text" \
+	--chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \
+	--save_steps="500" \
+	--eval_steps="500" \
+	--logging_steps="100" \
+	--layerdrop="0.0" \
+	--activation_dropout="0.1" \
+	--save_total_limit="3" \
+	--freeze_feature_encoder \
+	--feat_proj_dropout="0.0" \
+	--mask_time_prob="0.75" \
+	--mask_time_length="10" \
+	--mask_feature_prob="0.25" \
+	--mask_feature_length="64" \
+	--gradient_checkpointing \
+	--use_auth_token \
+	--fp16 \
+	--group_by_length \
+	--do_train --do_eval \
+	--push_to_hub

run_speech_recognition_ctc.py

@@ -0,0 +1,737 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# 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
+
+""" Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
+
+import functools
+import json
+import logging
+import os
+import re
+import sys
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+
+import datasets
+import numpy as np
+import torch
+from datasets import DatasetDict, load_dataset, load_metric
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    Wav2Vec2Processor,
+    set_seed,
+)
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.17.0.dev0")
+
+require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
+
+
+logger = logging.getLogger(__name__)
+
+
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    activation_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
+    )
+    feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
+    hidden_dropout: float = field(
+        default=0.0,
+        metadata={
+            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.05,
+        metadata={
+            "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
+            "vectors will be masked along the time axis."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+
+    Using `HfArgumentParser` we can turn this class
+    into argparse arguments to be able to specify them on
+    the command line.
+    """
+
+    dataset_name: str = field(
+        metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: str = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'test'"
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
+            "value if set."
+        },
+    )
+    chars_to_ignore: Optional[List[str]] = list_field(
+        default=None,
+        metadata={"help": "A list of characters to remove from the transcripts."},
+    )
+    eval_metrics: List[str] = list_field(
+        default=["wer"],
+        metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+            "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+            "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+            "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If :obj:`True`, will use the token generated when running"
+            ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
+        },
+    )
+    unk_token: str = field(
+        default="[UNK]",
+        metadata={"help": "The unk token for the tokenizer"},
+    )
+    pad_token: str = field(
+        default="[PAD]",
+        metadata={"help": "The padding token for the tokenizer"},
+    )
+    word_delimiter_token: str = field(
+        default="|",
+        metadata={"help": "The word delimiter token for the tokenizer"},
+    )
+    phoneme_language: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The target language that should be used be"
+            " passed to the tokenizer for tokenization. Note that"
+            " this is only relevant if the model classifies the"
+            " input audio to a sequence of phoneme sequences."
+        },
+    )
+
+
+@dataclass
+class DataCollatorCTCWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor (:class:`~transformers.AutoProcessor`)
+            The processor used for proccessing the data.
+        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        max_length (:obj:`int`, `optional`):
+            Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
+        max_length_labels (:obj:`int`, `optional`):
+            Maximum length of the ``labels`` returned list and optionally padding length (see above).
+        pad_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the sequence to a multiple of the provided value.
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+    """
+
+    processor: AutoProcessor
+    padding: Union[bool, str] = "longest"
+    pad_to_multiple_of: Optional[int] = None
+    pad_to_multiple_of_labels: Optional[int] = None
+
+    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lenghts and need
+        # different padding methods
+        input_features = [{"input_values": feature["input_values"]} for feature in features]
+        label_features = [{"input_ids": feature["labels"]} for feature in features]
+
+        batch = self.processor.pad(
+            input_features,
+            padding=self.padding,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors="pt",
+        )
+
+        with self.processor.as_target_processor():
+            labels_batch = self.processor.pad(
+                label_features,
+                padding=self.padding,
+                pad_to_multiple_of=self.pad_to_multiple_of_labels,
+                return_tensors="pt",
+            )
+
+        # replace padding with -100 to ignore loss correctly
+        labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def create_vocabulary_from_data(
+    datasets: DatasetDict,
+    word_delimiter_token: Optional[str] = None,
+    unk_token: Optional[str] = None,
+    pad_token: Optional[str] = None,
+):
+    # Given training and test labels create vocabulary
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target_text"])
+        vocab = list(set(all_text))
+        return {"vocab": [vocab], "all_text": [all_text]}
+
+    vocabs = datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets["train"].column_names,
+    )
+
+    # take union of all unique characters in each dataset
+    vocab_set = functools.reduce(
+        lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
+    )
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
+
+    # replace white space with delimiter token
+    if word_delimiter_token is not None:
+        vocab_dict[word_delimiter_token] = vocab_dict[" "]
+        del vocab_dict[" "]
+
+    # add unk and pad token
+    if unk_token is not None:
+        vocab_dict[unk_token] = len(vocab_dict)
+
+    if pad_token is not None:
+        vocab_dict[pad_token] = len(vocab_dict)
+
+    return vocab_dict
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Detecting last checkpoint.
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 1. First, let's load the dataset
+    raw_datasets = DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.audio_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--audio_column_name` to the correct audio column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.text_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--text_column_name` to the correct text column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.max_train_samples is not None:
+            raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.max_eval_samples is not None:
+            raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
+
+    # 2. We remove some special characters from the datasets
+    # that make training complicated and do not help in transcribing the speech
+    # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
+    # that could be easily picked up by the model
+    chars_to_ignore_regex = (
+        f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
+    )
+    text_column_name = data_args.text_column_name
+
+    def remove_special_characters(batch):
+        if chars_to_ignore_regex is not None:
+            batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
+        else:
+            batch["target_text"] = batch[text_column_name].lower() + " "
+        return batch
+
+    with training_args.main_process_first(desc="dataset map special characters removal"):
+        raw_datasets = raw_datasets.map(
+            remove_special_characters,
+            remove_columns=[text_column_name],
+            desc="remove special characters from datasets",
+        )
+
+    # save special tokens for tokenizer
+    word_delimiter_token = data_args.word_delimiter_token
+    unk_token = data_args.unk_token
+    pad_token = data_args.pad_token
+
+    # 3. Next, let's load the config as we might need it to create
+    # the tokenizer
+    # load config
+    config = AutoConfig.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # 4. Next, if no tokenizer file is defined,
+    # we create the vocabulary of the model by extracting all unique characters from
+    # the training and evaluation datasets
+    # We need to make sure that only first rank saves vocabulary
+    # make sure all processes wait until vocab is created
+    tokenizer_name_or_path = model_args.tokenizer_name_or_path
+    tokenizer_kwargs = {}
+    if tokenizer_name_or_path is None:
+        # save vocab in training output dir
+        tokenizer_name_or_path = training_args.output_dir
+
+        vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
+
+        with training_args.main_process_first():
+            if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
+                os.remove(vocab_file)
+
+        with training_args.main_process_first(desc="dataset map vocabulary creation"):
+            if not os.path.isfile(vocab_file):
+                os.makedirs(tokenizer_name_or_path, exist_ok=True)
+                vocab_dict = create_vocabulary_from_data(
+                    raw_datasets,
+                    word_delimiter_token=word_delimiter_token,
+                    unk_token=unk_token,
+                    pad_token=pad_token,
+                )
+
+                # save vocab dict to be loaded into tokenizer
+                with open(vocab_file, "w") as file:
+                    json.dump(vocab_dict, file)
+
+        # if tokenizer has just been created
+        # it is defined by `tokenizer_class` if present in config else by `model_type`
+        tokenizer_kwargs = {
+            "config": config if config.tokenizer_class is not None else None,
+            "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
+            "unk_token": unk_token,
+            "pad_token": pad_token,
+            "word_delimiter_token": word_delimiter_token,
+        }
+
+    # 5. Now we can instantiate the feature extractor, tokenizer and model
+    # Note for distributed training, the .from_pretrained methods guarantee that only
+    # one local process can concurrently download model & vocab.
+
+    # load feature_extractor and tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_name_or_path,
+        use_auth_token=data_args.use_auth_token,
+        **tokenizer_kwargs,
+    )
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # adapt config
+    config.update(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    # create model
+    model = AutoModelForCTC.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        config=config,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # freeze encoder
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
+
+    # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
+    # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
+    # so that we just need to set the correct target sampling rate and normalize the input
+    # via the `feature_extractor`
+
+    # make sure that dataset decodes audio with correct sampling rate
+    dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
+    if dataset_sampling_rate != feature_extractor.sampling_rate:
+        raw_datasets = raw_datasets.cast_column(
+            data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+        )
+
+    # derive max & min input length for sample rate & max duration
+    max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
+    min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
+    audio_column_name = data_args.audio_column_name
+    num_workers = data_args.preprocessing_num_workers
+
+    # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
+    phoneme_language = data_args.phoneme_language
+
+    # Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    def prepare_dataset(batch):
+        # load audio
+        sample = batch[audio_column_name]
+
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        batch["input_values"] = inputs.input_values[0]
+        batch["input_length"] = len(batch["input_values"])
+
+        # encode targets
+        additional_kwargs = {}
+        if phoneme_language is not None:
+            additional_kwargs["phonemizer_lang"] = phoneme_language
+
+        batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
+        return batch
+
+    with training_args.main_process_first(desc="dataset map preprocessing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=next(iter(raw_datasets.values())).column_names,
+            num_proc=num_workers,
+            desc="preprocess datasets",
+        )
+
+        def is_audio_in_length_range(length):
+            return length > min_input_length and length < max_input_length
+
+        # filter data that is shorter than min_input_length
+        vectorized_datasets = vectorized_datasets.filter(
+            is_audio_in_length_range,
+            num_proc=num_workers,
+            input_columns=["input_length"],
+        )
+
+    # 7. Next, we can prepare the training.
+    # Let's use word error rate (WER) as our evaluation metric,
+    # instantiate a data collator and the trainer
+
+    # Define evaluation metrics during training, *i.e.* word error rate, character error rate
+    eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
+
+    # for large datasets it is advised to run the preprocessing on a
+    # single machine first with ``args.preprocessing_only`` since there will mostly likely
+    # be a timeout when running the script in distributed mode.
+    # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
+    # cached dataset
+    if data_args.preprocessing_only:
+        logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
+        return
+
+    def compute_metrics(pred):
+        pred_logits = pred.predictions
+        pred_ids = np.argmax(pred_logits, axis=-1)
+
+        pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(pred_ids)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
+
+        metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
+
+        return metrics
+
+    # Now save everything to be able to create a single processor later
+    if is_main_process(training_args.local_rank):
+        # save feature extractor, tokenizer and config
+        feature_extractor.save_pretrained(training_args.output_dir)
+        tokenizer.save_pretrained(training_args.output_dir)
+        config.save_pretrained(training_args.output_dir)
+
+    try:
+        processor = AutoProcessor.from_pretrained(training_args.output_dir)
+    except (OSError, KeyError):
+        warnings.warn(
+            "Loading a processor from a feature extractor config that does not"
+            " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
+            " attribute to your `preprocessor_config.json` file to suppress this warning: "
+            " `'processor_class': 'Wav2Vec2Processor'`",
+            FutureWarning,
+        )
+        processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
+
+    # Instantiate custom data collator
+    data_collator = DataCollatorCTCWithPadding(processor=processor)
+
+    # Initialize Trainer
+    trainer = Trainer(
+        model=model,
+        data_collator=data_collator,
+        args=training_args,
+        compute_metrics=compute_metrics,
+        train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+    )
+
+    # 8. Finally, we can start training
+
+    # Training
+    if training_args.do_train:
+
+        # use last checkpoint if exist
+        if last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        elif os.path.isdir(model_args.model_name_or_path):
+            checkpoint = model_args.model_name_or_path
+        else:
+            checkpoint = None
+
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()
+
+        metrics = train_result.metrics
+        max_train_samples = (
+            data_args.max_train_samples
+            if data_args.max_train_samples is not None
+            else len(vectorized_datasets["train"])
+        )
+        metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
+
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate()
+        max_eval_samples = (
+            data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
+        )
+        metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Write model card and (optionally) push to hub
+    config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "speech-recognition",
+        "tags": ["automatic-speech-recognition", data_args.dataset_name],
+        "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
+        "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
+    }
+    if "common_voice" in data_args.dataset_name:
+        kwargs["language"] = config_name
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1de06e6749143de4b91ac63e2bc1f701d3282dbb2db4432d379ca611227854de
+size 2991

vocab.json

@@ -0,0 +1 @@
+{"가": 1, "각": 2, "간": 3, "갈": 4, "감": 5, "갑": 6, "값": 7, "갓": 8, "갔": 9, "강": 10, "갖": 11, "같": 12, "갚": 13, "개": 14, "객": 15, "갠": 16, "갤": 17, "갯": 18, "갱": 19, "거": 20, "걱": 21, "건": 22, "걷": 23, "걸": 24, "검": 25, "겁": 26, "것": 27, "겉": 28, "게": 29, "겐": 30, "겟": 31, "겠": 32, "겨": 33, "격": 34, "겪": 35, "견": 36, "결": 37, "겸": 38, "겹": 39, "겼": 40, "경": 41, "곁": 42, "계": 43, "고": 44, "곡": 45, "곤": 46, "곧": 47, "골": 48, "곰": 49, "곱": 50, "곳": 51, "공": 52, "곶": 53, "과": 54, "곽": 55, "관": 56, "괄": 57, "괌": 58, "광": 59, "괘": 60, "괭": 61, "괴": 62, "굉": 63, "교": 64, "구": 65, "국": 66, "군": 67, "굳": 68, "굴": 69, "굵": 70, "굶": 71, "굽": 72, "굿": 73, "궁": 74, "궈": 75, "권": 76, "궐": 77, "궤": 78, "귀": 79, "규": 80, "균": 81, "그": 82, "극": 83, "근": 84, "글": 85, "금": 86, "급": 87, "긋": 88, "긍": 89, "기": 90, "긴": 91, "길": 92, "김": 93, "깁": 94, "깃": 95, "깊": 96, "까": 97, "깎": 98, "깔": 99, "깝": 100, "깡": 101, "깥": 102, "깨": 103, "깬": 104, "꺼": 105, "껍": 106, "껏": 107, "껑": 108, "께": 109, "껴": 110, "꼈": 111, "꼬": 112, "꼭": 113, "꼴": 114, "꼼": 115, "꼽": 116, "꽁": 117, "꽂": 118, "꽃": 119, "꽝": 120, "꽥": 121, "꾸": 122, "꾼": 123, "꿀": 124, "꿇": 125, "꿈": 126, "꿎": 127, "꿔": 128, "꿨": 129, "꿰": 130, "뀌": 131, "뀐": 132, "끄": 133, "끅": 134, "끈": 135, "끊": 136, "끌": 137, "끓": 138, "끔": 139, "끗": 140, "끝": 141, "끼": 142, "끽": 143, "낀": 144, "낌": 145, "나": 146, "낙": 147, "난": 148, "날": 149, "낡": 150, "남": 151, "납": 152, "낫": 153, "났": 154, "낭": 155, "낮": 156, "낯": 157, "낳": 158, "내": 159, "낸": 160, "낼": 161, "냄": 162, "냅": 163, "냇": 164, "냈": 165, "냉": 166, "냐": 167, "냥": 168, "너": 169, "넉": 170, "넌": 171, "널": 172, "넓": 173, "넘": 174, "넛": 175, "넣": 176, "네": 177, "넥": 178, "넨": 179, "넷": 180, "녀": 181, "녁": 182, "년": 183, "념": 184, "녔": 185, "녕": 186, "녘": 187, "노": 188, "녹": 189, "논": 190, "놀": 191, "놈": 192, "농": 193, "높": 194, "놓": 195, "놨": 196, "뇌": 197, "뇨": 198, "뇽": 199, "누": 200, "눈": 201, "눌": 202, "눠": 203, "뉜": 204, "뉴": 205, "늉": 206, "느": 207, "는": 208, "늘": 209, "늠": 210, "능": 211, "늦": 212, "늪": 213, "늬": 214, "니": 215, "닉": 216, "닌": 217, "닐": 218, "님": 219, "닙": 220, "닛": 221, "닝": 222, "다": 223, "닥": 224, "닦": 225, "단": 226, "닫": 227, "달": 228, "닭": 229, "닮": 230, "닳": 231, "담": 232, "답": 233, "닷": 234, "당": 235, "닿": 236, "대": 237, "댁": 238, "댐": 239, "댓": 240, "더": 241, "덕": 242, "던": 243, "덜": 244, "덟": 245, "덤": 246, "덧": 247, "덩": 248, "덫": 249, "덮": 250, "데": 251, "덴": 252, "델": 253, "뎅": 254, "뎌": 255, "도": 256, "독": 257, "돈": 258, "돋": 259, "돌": 260, "돔": 261, "돕": 262, "돗": 263, "동": 264, "돛": 265, "돼": 266, "됐": 267, "되": 268, "된": 269, "될": 270, "됨": 271, "됩": 272, "두": 273, "둑": 274, "둔": 275, "둘": 276, "둠": 277, "둥": 278, "둬": 279, "뒀": 280, "뒤": 281, "뒷": 282, "듀": 283, "드": 284, "득": 285, "든": 286, "듣": 287, "들": 288, "듬": 289, "듭": 290, "듯": 291, "등": 292, "디": 293, "딘": 294, "딜": 295, "딥": 296, "딨": 297, "딩": 298, "딪": 299, "따": 300, "딱": 301, "딴": 302, "딸": 303, "땀": 304, "땄": 305, "땅": 306, "때": 307, "땐": 308, "땠": 309, "땡": 310, "떠": 311, "떡": 312, "떤": 313, "떨": 314, "떴": 315, "떻": 316, "떼": 317, "뗄": 318, "또": 319, "똑": 320, "똥": 321, "뚜": 322, "뚝": 323, "뚫": 324, "뛰": 325, "뛴": 326, "뜨": 327, "뜯": 328, "뜸": 329, "뜻": 330, "띄": 331, "띔": 332, "띠": 333, "띤": 334, "띨": 335, "띵": 336, "라": 337, "락": 338, "란": 339, "랄": 340, "람": 341, "랍": 342, "랐": 343, "랑": 344, "래": 345, "랙": 346, "랜": 347, "램": 348, "랫": 349, "랬": 350, "랭": 351, "랴": 352, "략": 353, "량": 354, "러": 355, "럭": 356, "런": 357, "럴": 358, "럼": 359, "럽": 360, "렀": 361, "렁": 362, "렇": 363, "레": 364, "렉": 365, "렌": 366, "렘": 367, "렛": 368, "려": 369, "력": 370, "련": 371, "렬": 372, "렴": 373, "렵": 374, "렷": 375, "렸": 376, "령": 377, "례": 378, "로": 379, "록": 380, "론": 381, "롤": 382, "롬": 383, "롭": 384, "롯": 385, "롱": 386, "뢰": 387, "료": 388, "룡": 389, "루": 390, "룬": 391, "룰": 392, "룸": 393, "룹": 394, "뤄": 395, "뤘": 396, "뤼": 397, "류": 398, "륙": 399, "륜": 400, "률": 401, "륭": 402, "르": 403, "륵": 404, "른": 405, "를": 406, "름": 407, "릅": 408, "릇": 409, "릉": 410, "릎": 411, "리": 412, "릭": 413, "린": 414, "릴": 415, "림": 416, "립": 417, "릿": 418, "링": 419, "마": 420, "막": 421, "만": 422, "많": 423, "맏": 424, "말": 425, "맑": 426, "맘": 427, "맙": 428, "맛": 429, "망": 430, "맞": 431, "맡": 432, "매": 433, "맥": 434, "맨": 435, "맹": 436, "맺": 437, "머": 438, "먹": 439, "먼": 440, "멀": 441, "멈": 442, "멋": 443, "멍": 444, "메": 445, "멕": 446, "멘": 447, "멜": 448, "멧": 449, "며": 450, "면": 451, "멸": 452, "명": 453, "몇": 454, "모": 455, "목": 456, "몫": 457, "몬": 458, "몰": 459, "몸": 460, "못": 461, "몽": 462, "묘": 463, "무": 464, "묵": 465, "묶": 466, "문": 467, "묻": 468, "물": 469, "뭄": 470, "뭇": 471, "뭐": 472, "뭔": 473, "뭘": 474, "뮤": 475, "뮬": 476, "므": 477, "믈": 478, "미": 479, "믹": 480, "민": 481, "믿": 482, "밀": 483, "밋": 484, "밌": 485, "밍": 486, "및": 487, "밑": 488, "바": 489, "박": 490, "밖": 491, "반": 492, "받": 493, "발": 494, "밝": 495, "밤": 496, "밥": 497, "방": 498, "밭": 499, "배": 500, "백": 501, "밴": 502, "뱀": 503, "뱃": 504, "뱅": 505, "버": 506, "벅": 507, "번": 508, "벌": 509, "범": 510, "법": 511, "벗": 512, "벚": 513, "베": 514, "벤": 515, "벨": 516, "벳": 517, "벼": 518, "벽": 519, "변": 520, "별": 521, "볍": 522, "볐": 523, "병": 524, "볕": 525, "보": 526, "복": 527, "볶": 528, "본": 529, "볼": 530, "봄": 531, "봅": 532, "봇": 533, "봉": 534, "봐": 535, "봤": 536, "뵈": 537, "뵙": 538, "부": 539, "북": 540, "분": 541, "불": 542, "붉": 543, "붐": 544, "붓": 545, "붕": 546, "붙": 547, "뷔": 548, "뷰": 549, "브": 550, "븐": 551, "블": 552, "비": 553, "빅": 554, "빈": 555, "빌": 556, "빗": 557, "빙": 558, "빚": 559, "빛": 560, "빠": 561, "빨": 562, "빵": 563, "빼": 564, "뺀": 565, "뺌": 566, "뺏": 567, "뺑": 568, "뻐": 569, "뻑": 570, "뻔": 571, "뻗": 572, "뻘": 573, "뼈": 574, "뽀": 575, "뽑": 576, "뽕": 577, "뿌": 578, "뿐": 579, "뿜": 580, "쁘": 581, "쁜": 582, "쁩": 583, "삐": 584, "사": 585, "삭": 586, "산": 587, "살": 588, "삶": 589, "삼": 590, "삽": 591, "삿": 592, "샀": 593, "상": 594, "새": 595, "색": 596, "샌": 597, "샐": 598, "샘": 599, "생": 600, "샤": 601, "샬": 602, "샵": 603, "샷": 604, "서": 605, "석": 606, "섞": 607, "선": 608, "섣": 609, "설": 610, "섬": 611, "섭": 612, "섯": 613, "섰": 614, "성": 615, "세": 616, "섹": 617, "센": 618, "셀": 619, "셈": 620, "셉": 621, "셋": 622, "셔": 623, "션": 624, "셜": 625, "셨": 626, "셰": 627, "소": 628, "속": 629, "손": 630, "솔": 631, "솜": 632, "솟": 633, "송": 634, "솥": 635, "쇄": 636, "쇠": 637, "쇤": 638, "쇼": 639, "숍": 640, "수": 641, "숙": 642, "순": 643, "술": 644, "숨": 645, "숭": 646, "숲": 647, "쉬": 648, "쉰": 649, "쉼": 650, "쉽": 651, "슈": 652, "슐": 653, "스": 654, "슨": 655, "슬": 656, "슴": 657, "습": 658, "슷": 659, "승": 660, "시": 661, "식": 662, "신": 663, "실": 664, "싫": 665, "심": 666, "십": 667, "싱": 668, "싶": 669, "싸": 670, "싹": 671, "싼": 672, "쌀": 673, "쌈": 674, "쌌": 675, "쌍": 676, "쌓": 677, "써": 678, "썩": 679, "썰": 680, "썼": 681, "쏘": 682, "쏜": 683, "쏟": 684, "쏠": 685, "쑤": 686, "쓰": 687, "쓴": 688, "쓸": 689, "씀": 690, "씁": 691, "씌": 692, "씨": 693, "씩": 694, "씬": 695, "씸": 696, "씻": 697, "아": 698, "악": 699, "안": 700, "앉": 701, "않": 702, "알": 703, "앓": 704, "암": 705, "압": 706, "앗": 707, "았": 708, "앙": 709, "앞": 710, "애": 711, "액": 712, "앤": 713, "앨": 714, "앱": 715, "앵": 716, "야": 717, "약": 718, "얇": 719, "양": 720, "얗": 721, "얘": 722, "어": 723, "억": 724, "언": 725, "얹": 726, "얻": 727, "얼": 728, "얽": 729, "엄": 730, "업": 731, "없": 732, "엇": 733, "었": 734, "엉": 735, "엎": 736, "에": 737, "엑": 738, "엔": 739, "엘": 740, "엠": 741, "엡": 742, "엣": 743, "여": 744, "역": 745, "연": 746, "열": 747, "엷": 748, "염": 749, "엽": 750, "엿": 751, "였": 752, "영": 753, "옆": 754, "예": 755, "옛": 756, "오": 757, "옥": 758, "온": 759, "올": 760, "옮": 761, "옳": 762, "옴": 763, "옵": 764, "옷": 765, "옹": 766, "와": 767, "완": 768, "왈": 769, "왔": 770, "왕": 771, "왜": 772, "외": 773, "왼": 774, "요": 775, "욕": 776, "용": 777, "우": 778, "욱": 779, "운": 780, "울": 781, "움": 782, "웁": 783, "웃": 784, "웅": 785, "워": 786, "원": 787, "월": 788, "웠": 789, "웨": 790, "웬": 791, "웹": 792, "위": 793, "윈": 794, "윌": 795, "윗": 796, "윙": 797, "유": 798, "육": 799, "윤": 800, "율": 801, "융": 802, "으": 803, "은": 804, "을": 805, "음": 806, "읍": 807, "응": 808, "의": 809, "이": 810, "익": 811, "인": 812, "일": 813, "읽": 814, "잃": 815, "임": 816, "입": 817, "잇": 818, "있": 819, "잉": 820, "잊": 821, "잎": 822, "자": 823, "작": 824, "잔": 825, "잖": 826, "잘": 827, "잠": 828, "잡": 829, "잣": 830, "장": 831, "잦": 832, "재": 833, "잭": 834, "잰": 835, "잽": 836, "쟁": 837, "저": 838, "적": 839, "전": 840, "절": 841, "젊": 842, "젋": 843, "점": 844, "접": 845, "젓": 846, "정": 847, "젖": 848, "제": 849, "젝": 850, "젠": 851, "젤": 852, "져": 853, "젼": 854, "졌": 855, "조": 856, "족": 857, "존": 858, "졸": 859, "좀": 860, "좁": 861, "종": 862, "좋": 863, "좌": 864, "죄": 865, "죠": 866, "주": 867, "죽": 868, "준": 869, "줄": 870, "줍": 871, "중": 872, "줘": 873, "줬": 874, "쥐": 875, "쥔": 876, "쥘": 877, "쥬": 878, "즈": 879, "즉": 880, "즌": 881, "즐": 882, "즘": 883, "증": 884, "지": 885, "직": 886, "진": 887, "질": 888, "짊": 889, "짐": 890, "집": 891, "짓": 892, "징": 893, "짖": 894, "짙": 895, "짚": 896, "짜": 897, "짝": 898, "짧": 899, "짬": 900, "째": 901, "쨌": 902, "쩌": 903, "쩍": 904, "쩔": 905, "쩜": 906, "쪼": 907, "쪽": 908, "쫄": 909, "쫓": 910, "쭉": 911, "쯤": 912, "찌": 913, "찍": 914, "찔": 915, "찢": 916, "찧": 917, "차": 918, "착": 919, "찬": 920, "찮": 921, "찰": 922, "참": 923, "찻": 924, "창": 925, "찾": 926, "채": 927, "책": 928, "챌": 929, "챔": 930, "챙": 931, "챠": 932, "처": 933, "척": 934, "천": 935, "철": 936, "첨": 937, "첩": 938, "첫": 939, "청": 940, "체": 941, "첸": 942, "첼": 943, "쳇": 944, "쳐": 945, "쳤": 946, "초": 947, "촉": 948, "촌": 949, "촘": 950, "촛": 951, "총": 952, "촨": 953, "촬": 954, "최": 955, "추": 956, "축": 957, "춘": 958, "출": 959, "춤": 960, "충": 961, "춰": 962, "취": 963, "츠": 964, "측": 965, "츰": 966, "층": 967, "치": 968, "칙": 969, "친": 970, "칠": 971, "침": 972, "칩": 973, "칫": 974, "칭": 975, "카": 976, "칸": 977, "칼": 978, "캉": 979, "캐": 980, "캔": 981, "캘": 982, "캠": 983, "커": 984, "컥": 985, "컨": 986, "컫": 987, "컴": 988, "컵": 989, "컷": 990, "컸": 991, "케": 992, "켈": 993, "켐": 994, "켑": 995, "켓": 996, "켜": 997, "켰": 998, "코": 999, "콘": 1000, "콜": 1001, "콤": 1002, "콥": 1003, "콧": 1004, "콩": 1005, "쾌": 1006, "쿄": 1007, "쿠": 1008, "쿡": 1009, "쿨": 1010, "쿼": 1011, "퀴": 1012, "큐": 1013, "크": 1014, "큰": 1015, "클": 1016, "큼": 1017, "키": 1018, "킥": 1019, "킨": 1020, "킬": 1021, "킷": 1022, "킹": 1023, "타": 1024, "탁": 1025, "탄": 1026, "탈": 1027, "탐": 1028, "탑": 1029, "탓": 1030, "탕": 1031, "태": 1032, "택": 1033, "탠": 1034, "탬": 1035, "탱": 1036, "터": 1037, "턱": 1038, "턴": 1039, "털": 1040, "텃": 1041, "텅": 1042, "테": 1043, "텍": 1044, "텐": 1045, "텔": 1046, "템": 1047, "텼": 1048, "토": 1049, "톡": 1050, "톤": 1051, "톨": 1052, "톰": 1053, "통": 1054, "퇴": 1055, "투": 1056, "툴": 1057, "툼": 1058, "퉁": 1059, "튀": 1060, "튜": 1061, "튬": 1062, "트": 1063, "특": 1064, "튼": 1065, "튿": 1066, "틀": 1067, "틈": 1068, "티": 1069, "틱": 1070, "틴": 1071, "틸": 1072, "팀": 1073, "팅": 1074, "파": 1075, "팍": 1076, "팎": 1077, "판": 1078, "팔": 1079, "팜": 1080, "팡": 1081, "패": 1082, "팩": 1083, "팬": 1084, "팰": 1085, "팻": 1086, "팽": 1087, "퍼": 1088, "펀": 1089, "펄": 1090, "펌": 1091, "페": 1092, "펜": 1093, "펠": 1094, "펫": 1095, "펴": 1096, "편": 1097, "펼": 1098, "폄": 1099, "폈": 1100, "평": 1101, "폐": 1102, "포": 1103, "폭": 1104, "폰": 1105, "폴": 1106, "폼": 1107, "표": 1108, "푸": 1109, "푹": 1110, "푼": 1111, "풀": 1112, "품": 1113, "풋": 1114, "풍": 1115, "퓨": 1116, "퓰": 1117, "프": 1118, "픈": 1119, "플": 1120, "픔": 1121, "피": 1122, "픽": 1123, "핀": 1124, "필": 1125, "핏": 1126, "핑": 1127, "하": 1128, "학": 1129, "한": 1130, "할": 1131, "함": 1132, "합": 1133, "핫": 1134, "항": 1135, "해": 1136, "핵": 1137, "핸": 1138, "햇": 1139, "했": 1140, "행": 1141, "향": 1142, "허": 1143, "헌": 1144, "헐": 1145, "험": 1146, "헝": 1147, "헤": 1148, "헨": 1149, "헬": 1150, "헴": 1151, "헷": 1152, "혀": 1153, "혁": 1154, "현": 1155, "혈": 1156, "혐": 1157, "협": 1158, "혔": 1159, "형": 1160, "혜": 1161, "호": 1162, "혹": 1163, "혼": 1164, "홀": 1165, "홈": 1166, "홉": 1167, "홍": 1168, "화": 1169, "확": 1170, "환": 1171, "활": 1172, "황": 1173, "회": 1174, "획": 1175, "횡": 1176, "효": 1177, "후": 1178, "훈": 1179, "훌": 1180, "훤": 1181, "훨": 1182, "훼": 1183, "휘": 1184, "휩": 1185, "휴": 1186, "흉": 1187, "흐": 1188, "흑": 1189, "흔": 1190, "흘": 1191, "흙": 1192, "흠": 1193, "흡": 1194, "흥": 1195, "흩": 1196, "희": 1197, "흰": 1198, "히": 1199, "힌": 1200, "힐": 1201, "힘": 1202, "|": 0, "[UNK]": 1203, "[PAD]": 1204}