first

.gitattributes

@@ -25,3 +25,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*unigram*.* filter=lfs diff=lfs merge=lfs -text
+*.txt filter=lfs diff=lfs merge=lfs -text

README.md

@@ -63,16 +63,16 @@ As you see from the results above, adding even a simple 5-gram language will imp
 ### Parameters
 The final model was run using these parameters:
 ```
---dataset_name="NbAiLab/NPSC"
---model_name_or_path="facebook/wav2vec2-xls-r-1b"
---dataset_config_name="16K_mp3_bokmaal"
---output_dir="./"
---overwrite_output_dir
---num_train_epochs="40"
---per_device_train_batch_size="12"
---per_device_eval_batch_size="12" 
+--dataset_name="NbAiLab/NPSC" 
+--model_name_or_path="KBLab/wav2vec2-large-voxrex" 
+--dataset_config_name="16K_mp3_bokmaal" 
+--output_dir="./" 
+--overwrite_output_dir 
+--num_train_epochs="15" 
+--per_device_train_batch_size="16" 
+--per_device_eval_batch_size="16" 
 --gradient_accumulation_steps="2" 
---learning_rate="2e-5" 
+--learning_rate="1e-4" 
 --warmup_steps="2000" 
 --length_column_name="input_length" 
 --evaluation_strategy="steps" 
@@ -84,24 +84,23 @@ The final model was run using these parameters:
 --attention_dropout="0.094" 
 --activation_dropout="0.055" 
 --hidden_dropout="0.047" 
---save_total_limit="3"
+--save_total_limit="3" 
 --freeze_feature_encoder 
 --feat_proj_dropout="0.04" 
 --mask_time_prob="0.082" 
 --mask_time_length="10" 
 --mask_feature_prob="0.25" 
 --mask_feature_length="64" 
---gradient_checkpointing
+--gradient_checkpointing 
 --min_duration_in_seconds="0.5" 
 --max_duration_in_seconds="30.0" 
-	--ctc_zero_infinity=True 
 --use_auth_token 
 --seed="42" 
 --fp16 
 --group_by_length 
 --do_train --do_eval 
 --push_to_hub 
---preprocessing_num_workers="16"
+--preprocessing_num_workers="32"
 ```
 
 Following this settings, the training might take 3-4 days on an average GPU. You should however get a decent model and faster results by tweaking these parameters

add_kenlm.py

@@ -0,0 +1,34 @@
+import argparse
+from transformers import AutoProcessor
+from transformers import Wav2Vec2ProcessorWithLM
+from pyctcdecode import build_ctcdecoder
+
+
+def main(args):
+    processor = AutoProcessor.from_pretrained(args.model_name_or_path)
+    vocab_dict = processor.tokenizer.get_vocab()
+    sorted_vocab_dict = {
+        k.lower(): v for k, v in sorted(vocab_dict.items(), key=lambda item: item[1])
+    }
+    decoder = build_ctcdecoder(
+        labels=list(sorted_vocab_dict.keys()),
+        kenlm_model_path=args.kenlm_model_path,
+    )
+    processor_with_lm = Wav2Vec2ProcessorWithLM(
+        feature_extractor=processor.feature_extractor,
+        tokenizer=processor.tokenizer,
+        decoder=decoder,
+    )
+    processor_with_lm.save_pretrained(args.model_name_or_path)
+    print(f"Run: ~/bin/build_binary language_model/*.arpa language_model/5gram.bin -T $(pwd) && rm language_model/*.arpa")
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model_name_or_path', default="./", help='Model name or path. Defaults to ./')
+    parser.add_argument('--kenlm_model_path', required=True, help='Path to KenLM arpa file.')
+    args = parser.parse_args()
+    return args
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

added_tokens.json

@@ -0,0 +1 @@
+{"<s>": 32, "</s>": 33}

all_results.json

@@ -0,0 +1,14 @@
+{
+    "epoch": 1.5,
+    "eval_loss": 0.12273009121417999,
+    "eval_runtime": 399.7312,
+    "eval_samples": 5437,
+    "eval_samples_per_second": 13.602,
+    "eval_steps_per_second": 0.851,
+    "eval_wer": 0.0989537882279114,
+    "train_loss": 0.0,
+    "train_runtime": 5.9011,
+    "train_samples": 49645,
+    "train_samples_per_second": 841.279,
+    "train_steps_per_second": 26.436
+}

alphabet.json

@@ -0,0 +1 @@
+{"labels": [" ", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "\u00e5", "\u00e6", "\u00f8", "\u2047", "", "<s>", "</s>"], "is_bpe": false}

config.json

@@ -0,0 +1,115 @@
+{
+  "_name_or_path": "./",
+  "activation_dropout": 0.055,
+  "adapter_kernel_size": 3,
+  "adapter_stride": 2,
+  "add_adapter": false,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForCTC"
+  ],
+  "attention_dropout": 0.094,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "codevector_dim": 768,
+  "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": true,
+  "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": 0.04,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.047,
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.041,
+  "mask_channel_length": 10,
+  "mask_channel_min_space": 1,
+  "mask_channel_other": 0.0,
+  "mask_channel_prob": 0.0,
+  "mask_channel_selection": "static",
+  "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_min_space": 1,
+  "mask_time_other": 0.0,
+  "mask_time_prob": 0.082,
+  "mask_time_selection": "static",
+  "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": 24,
+  "num_negatives": 100,
+  "output_hidden_size": 1024,
+  "pad_token_id": 31,
+  "proj_codevector_dim": 768,
+  "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": 34,
+  "xvector_output_dim": 512
+}

eval.py

@@ -0,0 +1,151 @@
+#!/usr/bin/env python3
+import argparse
+import re
+from typing import Dict
+
+import torch
+from datasets import Audio, Dataset, load_dataset, load_metric
+
+from transformers import AutoFeatureExtractor, pipeline
+
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """DO NOT CHANGE. This function computes and logs the result metrics."""
+
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+
+
+def normalize_text(text: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+
+    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'  # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
+
+    text = re.sub(chars_to_ignore_regex, "", text.lower()) + " "
+    text = re.sub('[áàâ]', 'a', text)
+    text = re.sub('[ä]', 'æ', text)
+    text = re.sub('[éèëê]', 'e', text)
+    text = re.sub('[íìïî]', 'i', text)
+    text = re.sub('[óòöô]', 'o', text)
+    text = re.sub('[ö]', 'ø', text)
+    text = re.sub('[ç]', 'c', text)
+    text = re.sub('[úùüû]', 'u', text)
+    text = re.sub('\s', ' ', text)
+    text = re.sub('<ee>', 'eee', text)
+    text = re.sub('<qq>', 'qqq', text)
+    text = re.sub('<mm>', 'mmm', text)
+    text = re.sub('<inaudible>', 'xxx', text)
+    text = re.sub('[<>]', '', text)
+
+    # # In addition, we can normalize the target text, e.g. removing new lines characters etc...
+    # # note that order is important here!
+    # token_sequences_to_ignore = ["\n\n", "\n", "   ", "  "]
+
+    # for t in token_sequences_to_ignore:
+    #     text = " ".join(text.split(t))
+
+    return text
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+
+    # for testing: only process the first two examples as a test
+    # dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    if args.device is None:
+        args.device = 0 if torch.cuda.is_available() else -1
+    asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(
+            batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
+        )
+
+        batch["prediction"] = prediction["text"]
+        batch["target"] = normalize_text(batch["text"])
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
+    )
+    parser.add_argument(
+        "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=None,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    args = parser.parse_args()
+
+    main(args)

eval_results.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 1.5,
+    "eval_loss": 0.12273009121417999,
+    "eval_runtime": 399.7312,
+    "eval_samples": 5437,
+    "eval_samples_per_second": 13.602,
+    "eval_steps_per_second": 0.851,
+    "eval_wer": 0.0989537882279114
+}

grid.csv

@@ -0,0 +1,101 @@
+alpha,beta,wer,cer
+0.001,0.001,0.09547085935319984,0.029674778514815793
+0.001,0.01,0.09553821432351554,0.029683084728361897
+0.001,0.1,0.09571141281861306,0.029724615796092424
+0.001,0.25,0.096019321254342,0.029746211951312298
+0.001,0.5,0.0972509549972577,0.029950544804546488
+0.001,0.75,0.09849221087878993,0.030179796298418996
+0.001,1,0.09977195531478826,0.030379145423525523
+0.001,1.5,0.1028799060879271,0.03088914693525639
+0.001,2,0.10715213563366593,0.031522080407469615
+0.001,3,0.1189200111616808,0.03319495181565522
+0.01,0.001,0.09309419111491721,0.02945051074907095
+0.01,0.01,0.09311343539215026,0.029443865778234067
+0.01,0.1,0.09344058810511224,0.02948373560325537
+0.01,0.25,0.09395056145178828,0.029523605428276676
+0.01,0.5,0.09508597380853868,0.02965816608772358
+0.01,0.75,0.0962887411356048,0.029847547756574784
+0.01,1,0.09774168406670067,0.030105040376504046
+0.01,1.5,0.1008496348398395,0.030596768218433483
+0.01,2,0.10507375369249569,0.031251297845866575
+0.01,3,0.11661069789371385,0.03288928315715854
+0.1,0.001,0.08018128109153541,0.027892265087821597
+0.1,0.01,0.08017165895291888,0.027895587573240038
+0.1,0.1,0.0802775024777007,0.027907216272204584
+0.1,0.25,0.0806142773292792,0.027893926330530817
+0.1,0.5,0.08166309043848086,0.028045099417069935
+0.1,0.75,0.08281774707246432,0.028191288775481386
+0.1,1,0.08378958307273375,0.02827435091094244
+0.1,1.5,0.08655113685567754,0.028596631996531324
+0.1,2,0.08970719832189902,0.02901194267383659
+0.1,3,0.09869427578973702,0.03040738654958228
+0.25,0.001,0.07359973827782963,0.027008483966515992
+0.25,0.01,0.07358049400059657,0.027008483966515992
+0.25,0.1,0.07362860469367921,0.026993532782133
+0.25,0.25,0.07380180318877674,0.026963630413367023
+0.25,0.5,0.07414820017897178,0.02695698544253014
+0.25,0.75,0.07442724219885112,0.02695698544253014
+0.25,1,0.07499494837722633,0.027063304975920285
+0.25,1.5,0.07606300576366103,0.027172946994728876
+0.25,2,0.07761217008092218,0.027242719188516163
+0.25,3,0.08315452192404284,0.02800689083475785
+0.5,0.001,0.07044367681160815,0.026963630413367023
+0.5,0.01,0.0704244325343751,0.02695532419982092
+0.5,0.1,0.07028972259374369,0.026870600821650645
+0.5,0.25,0.07050140964330733,0.02688222952061519
+0.5,0.5,0.07065536386117179,0.026905486918544285
+0.5,0.75,0.07108836009891559,0.02694037301543793
+0.5,1,0.07091516160381807,0.026809134841409465
+0.5,1.5,0.0711942036236974,0.026624736900685928
+0.5,2,0.07201208540610236,0.026729395191366852
+0.5,3,0.0740904673472726,0.026917115617508834
+0.75,0.001,0.07321485273316847,0.02827767339636088
+0.75,0.01,0.07318598631731889,0.028264383454687115
+0.75,0.1,0.07289732215882302,0.028197933746318272
+0.75,0.25,0.07285883360435691,0.028173015105679954
+0.75,0.5,0.07246432592107922,0.02800855207746707
+0.75,0.75,0.07206981823780154,0.027877313903438606
+0.75,1,0.07177153194068914,0.027717834603353385
+0.75,1.5,0.07124231431678005,0.02735069996461553
+0.75,2,0.07131929142571228,0.027231090489551614
+0.75,3,0.07209868465365112,0.027249364159353046
+1,0.001,0.08050843380449739,0.030586800762178155
+1,0.01,0.08048918952726433,0.030598429461142704
+1,0.1,0.08019090323015193,0.03053197975277386
+1,0.25,0.07958470849731061,0.03034758181205032
+1,0.5,0.07864173891289078,0.030065170551482744
+1,0.75,0.07811252128898169,0.029845886513865563
+1,1,0.07759292580368912,0.029613312534574613
+1,1.5,0.07614960501120979,0.029111617236389855
+1,2,0.07533172322880484,0.0288076098206024
+1,3,0.07439837578300153,0.028271028425523998
+1.5,0.001,0.10399607416744445,0.035783167956621634
+1.5,0.01,0.10393834133574528,0.03576987801494786
+1.5,0.1,0.10341874585045273,0.0356070762294442
+1.5,0.25,0.10243728771156678,0.035369518522025585
+1.5,0.5,0.10124414252311718,0.035130299571897755
+1.5,0.75,0.1000798637505172,0.03490270932073447
+1.5,1,0.09854032157187256,0.034550525866379606
+1.5,1.5,0.09586536703647752,0.03397573588898912
+1.5,2,0.09308456897630067,0.03324811158235029
+1.5,3,0.08891818295534365,0.032244720985980774
+2,0.001,0.12878270324362293,0.040386471503873186
+2,0.01,0.1287730811050064,0.04037650404761786
+2,0.1,0.12826310775833036,0.04026520078610005
+2,0.25,0.12718542823327914,0.040067512903702744
+2,0.5,0.1253187333416725,0.03970037826496489
+2,0.75,0.1241833209849221,0.03952096405236902
+2,1,0.12279773302414194,0.03930500250017028
+2,1.5,0.11951658375590558,0.0386836977269216
+2,2,0.11614883524012047,0.03804578052658071
+2,3,0.11092401397134527,0.03697760146455157
+3,0.001,0.15472398895378486,0.04477547474163523
+3,0.01,0.15461814542900304,0.04474889485828769
+3,0.1,0.15435834768635676,0.04472895994577704
+3,0.25,0.15378101936936503,0.044596060529039354
+3,0.5,0.1530016261414262,0.0444498711706279
+3,0.75,0.15221261077487083,0.04431863299659944
+3,1,0.15129850760630056,0.04415416996838655
+3,1.5,0.14928748063544603,0.04384351758176221
+3,2,0.14715136586257663,0.04347472170031514
+3,3,0.14315817833671712,0.042783644733279176

grid.py

@@ -0,0 +1,34 @@
+#!/usr/bin/env python
+from argparse import ArgumentDefaultsHelpFormatter
+from collections import namedtuple
+from functools import partialmethod
+import json
+
+from tqdm import tqdm
+from eval import main
+
+
+# tqdm.__init__ = partialmethod(tqdm.__init__, disable=True)
+
+Args = namedtuple("Args", "model_id dataset config split log_outputs chunk_length_s stride_length_s device")
+args = Args("./", "NbAiLab/NPSC", "16K_mp3_bokmaal", "test", True, None, None, 0)
+with open("grid.csv", "w") as grid:
+    grid.write("alpha,beta,wer,cer")
+    for alpha in [0.001, 0.01, 0.1, 0.25, 0.5, 0.75, 1, 1.5, 2, 3]:
+        for beta in [0.001, 0.01, 0.1, 0.25, 0.5, 0.75, 1, 1.5, 2, 3]:
+            with open("./language_model/attrs.json", "r") as attrs_file:
+                attrs = json.load(attrs_file)
+            attrs["alpha"] = alpha
+            attrs["beta"] = beta
+            with open("./language_model/attrs.json", "w") as attrs_file:
+                json.dump(attrs, attrs_file)
+            print(f"alpha = {alpha}, beta = {beta}")
+            main(args)
+            with open("NbAiLab_NPSC_16K_mp3_bokmaal_test_eval_results.txt") as results_file:
+                results = results_file.read().strip().split("\n")
+            wer = float(results[0][5:])
+            cer = float(results[1][5:])
+            grid.write(f"\n{alpha},{beta},{wer},{cer}")
+            print("--------------")
+
+

language_model/5gram.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b41c24c63f2f0585bea83666369593f3b3e6d047f327a90f36ebca2c35ef0ff
+size 4243671427

language_model/attrs.json

@@ -0,0 +1 @@
+{"alpha": 0.5, "beta": 0.1, "unk_score_offset": -10.0, "score_boundary": true}

language_model/unigrams.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac3e71ca49838ca355df6fdcb8d89344a5a9bf9e1a76587cdf5df1367c19b9a9
+size 16759269

log_NbAiLab_NPSC_16K_mp3_bokmaal_test_eval_results.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1ee0d706882f015034b183fde35453b0c0c02f69e5143fb9fc329780fea0f27
+size 50

log_NbAiLab_NPSC_16K_mp3_bokmaal_test_predictions.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60928a28777980ce5e8abb4fed8c3b823135c1e5b5247c6b7dfdde0435440d08
+size 644626

log_NbAiLab_NPSC_16K_mp3_bokmaal_test_targets.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7e27a71ec76c0ded0b0cf852b9d42c620e1c5863bcf9a06fcbbbbd924e9d1b9
+size 654671

preprocessor_config.json

@@ -0,0 +1,10 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0,
+  "processor_class": "Wav2Vec2ProcessorWithLM",
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:146aae60aca5aad96c2d676908ba62300d18e34ec2cc31c921f29458eacbf2a7
+size 1262063089

run.sh

@@ -0,0 +1,39 @@
+python run_speech_recognition_ctc.py \
+        --dataset_name="NbAiLab/NPSC" \
+        --model_name_or_path="KBLab/wav2vec2-large-voxrex" \
+        --dataset_config_name="16K_mp3" \
+        --output_dir="./" \
+        --overwrite_output_dir \
+        --num_train_epochs="15" \
+        --per_device_train_batch_size="16" \
+        --per_device_eval_batch_size="16" \
+        --gradient_accumulation_steps="2" \
+        --learning_rate="1e-4" \
+        --warmup_steps="2000" \
+        --length_column_name="input_length" \
+        --evaluation_strategy="steps" \
+        --text_column_name="text" \
+        --save_steps="500" \
+        --eval_steps="500" \
+        --logging_steps="100" \
+        --layerdrop="0.041" \
+        --attention_dropout="0.094" \
+        --activation_dropout="0.055" \
+        --hidden_dropout="0.047" \
+        --save_total_limit="3" \
+        --freeze_feature_encoder \
+        --feat_proj_dropout="0.04" \
+        --mask_time_prob="0.082" \
+        --mask_time_length="10" \
+        --mask_feature_prob="0.25" \
+        --mask_feature_length="64" \
+        --gradient_checkpointing \
+        --min_duration_in_seconds="0.5" \
+        --max_duration_in_seconds="30.0" \
+        --use_auth_token \
+        --seed="42" \
+        --fp16 \
+        --group_by_length \
+        --do_train --do_eval \
+        --push_to_hub \
+        --preprocessing_num_workers="32"

run_speech_recognition_ctc.py

@@ -0,0 +1,792 @@
+#!/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("4.16.0.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'."}
+    )
+    ctc_zero_infinity: Optional[bool] = field(
+        default=False, metadata={"help": "If True, will try yo aboud the CTC loss goinf to infinity."}
+    )
+
+@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+validation",
+        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 'train'"
+        },
+    )
+    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)
+
+    # Pre-processing dataset
+    import re
+    def filter_numeric(entry):
+        return  (
+            "0" not in entry["text"]
+            and "1" not in entry["text"]
+            and "2" not in entry["text"]
+            and "3" not in entry["text"]
+            and "4" not in entry["text"]
+            and "5" not in entry["text"]
+            and "6" not in entry["text"]
+            and "7" not in entry["text"]
+            and "8" not in entry["text"]
+            and "9" not in entry["text"]
+        )
+
+    def filter_inaudible(entry):
+        return not re.search("\d|<inaudible>", entry["text"], flags=re.IGNORECASE)
+
+    #def filter_nynorsk(entry):
+    #    return re.search("nb-no", entry["sentence_language_code"], flags=re.IGNORECASE)
+
+    def filter_tooshort(entry):
+        #print(f"The audio sample ({entry["audio"]["path"]}) is too small, and has been omitted. "
+        return (len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)
+
+    def map_dataset(entry):
+        batch = {"text": entry["text"].lower()}
+        batch["text"] = re.sub('[áàâ]', 'a', batch["text"])
+        batch["text"] = re.sub('[ä]', 'æ', batch["text"])
+        batch["text"] = re.sub('[éèëê]', 'e', batch["text"])
+        batch["text"] = re.sub('[íìïî]', 'i', batch["text"])
+        batch["text"] = re.sub('[óòöô]', 'o', batch["text"])
+        batch["text"] = re.sub('[ö]', 'ø', batch["text"])
+        batch["text"] = re.sub('[ç]', 'c', batch["text"])
+        batch["text"] = re.sub('[úùüû]', 'u', batch["text"])
+        batch["text"] = re.sub('\s', ' ', batch["text"])
+        batch["text"] = re.sub('<ee>', 'eee', batch["text"])
+        batch["text"] = re.sub('<qq>', 'qqq', batch["text"])
+        batch["text"] = re.sub('<mm>', 'mmm', batch["text"])
+        # batch["text"] = re.sub('<inaudible>', '?', batch["text"])
+        if "<" in batch["text"]:
+            raise ValueError(batch["text"])
+        return batch
+
+    # 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,
+        ).shuffle()
+        raw_datasets["train"] = raw_datasets["train"].filter(filter_numeric).filter(filter_inaudible).filter(filter_tooshort)
+        raw_datasets["train"] = raw_datasets["train"].map(map_dataset)
+
+        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,
+        ).shuffle()
+        raw_datasets["eval"] = raw_datasets["eval"].filter(filter_numeric).filter(filter_inaudible).filter(filter_tooshort)
+        raw_datasets["eval"] = raw_datasets["eval"].map(map_dataset)
+
+        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
+    #)
+    chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'
+    
+    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,
+            "ctc_zero_infinity": model_args.ctc_zero_infinity,
+            "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()

runs/Feb04_12-19-59_dante/1643973630.5333126/events.out.tfevents.1643973630.dante.1786377.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:102d4a9860ed99fb120d6b61b95f00e0e274e6481585e1e630824a6b75940f95
+size 4765

runs/Feb04_12-19-59_dante/events.out.tfevents.1643973630.dante.1786377.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48a422de8738b9c59fa5e4bc553d5b0167e3b63a95af435473a699f5b81aa178
+size 52574

runs/Feb05_11-55-40_dante/1644058575.319772/events.out.tfevents.1644058575.dante.2573362.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:956f52ac1957263e06ed215d081ab32f3b0e5a37d3abf6613fbf63b0e1e5d732
+size 4765

runs/Feb05_11-55-40_dante/events.out.tfevents.1644058575.dante.2573362.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ed0c086f730c38fc09d64f5ffafdcefd8d80c2a6cc2cf5adabbf7b7255345a3
+size 10165

runs/Feb05_11-55-40_dante/events.out.tfevents.1644066583.dante.2573362.2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5e49ff79ccc0ad21eb2833b98a8f46f3dd861968bd63986f8da5f15ddc0a81f
+size 358

runs/Feb06_00-33-18_dante/1644104059.4582894/events.out.tfevents.1644104059.dante.2706942.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:adc02f1aa0c8de2f18a4eb7567614d49955912b374acf7bb2d7d4402b22715a2
+size 4765

runs/Feb06_00-33-18_dante/events.out.tfevents.1644104059.dante.2706942.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be31c7c9940fea61350188c3c5852d842fea96dfcf4b1e3d73fef55dc677c878
+size 34596

runs/Feb06_12-59-06_dante/1644148807.511835/events.out.tfevents.1644148807.dante.2790531.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de29a2a4263b81525551cc6a8648135a9470f06fdbc742c6b46db893fe00ab19
+size 4765

runs/Feb06_12-59-06_dante/events.out.tfevents.1644148807.dante.2790531.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4bd6d012fbd254fed1b5d06c38715b576b2c2c5bb7dd25ae0428906ff3fa494
+size 5285

runs/Feb06_13-01-43_dante/1644148960.1271484/events.out.tfevents.1644148960.dante.2793694.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da37a286fbcdc1d8756d26fc45a4cc591b00979834d1e0ef7fecffb5d808b7ca
+size 4765

runs/Feb06_13-01-43_dante/events.out.tfevents.1644148960.dante.2793694.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc132a43de8b7d0a6133ee02cc3bdb5eb32925d411ed9f1bdc793156b35c14a4
+size 5285

runs/Feb06_13-01-43_dante/events.out.tfevents.1644149411.dante.2793694.2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9395b424df11af082ef53c2069f670a722ff315edcd97097a68909cfcf8b1cb9
+size 358

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}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"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", "processor_class": "Wav2Vec2ProcessorWithLM"}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 1.5,
+    "train_loss": 0.0,
+    "train_runtime": 5.9011,
+    "train_samples": 49645,
+    "train_samples_per_second": 841.279,
+    "train_steps_per_second": 26.436
+}

trainer_state.json

@@ -0,0 +1,217 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.5001611343860781,
+  "global_step": 2327,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.06,
+      "learning_rate": 8.01984817006747e-06,
+      "loss": 0.2478,
+      "step": 100
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 7.659729258075218e-06,
+      "loss": 0.222,
+      "step": 200
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 7.303211535202886e-06,
+      "loss": 0.283,
+      "step": 300
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 6.943092623210635e-06,
+      "loss": 0.2102,
+      "step": 400
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 6.5829737112183816e-06,
+      "loss": 0.2127,
+      "step": 500
+    },
+    {
+      "epoch": 0.32,
+      "eval_loss": 0.13347241282463074,
+      "eval_runtime": 385.6175,
+      "eval_samples_per_second": 14.099,
+      "eval_steps_per_second": 0.882,
+      "eval_wer": 0.1047094237500858,
+      "step": 500
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 6.222854799226128e-06,
+      "loss": 0.1923,
+      "step": 600
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 5.862735887233875e-06,
+      "loss": 0.2001,
+      "step": 700
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 5.502616975241623e-06,
+      "loss": 0.199,
+      "step": 800
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 5.14249806324937e-06,
+      "loss": 0.1944,
+      "step": 900
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 4.782379151257117e-06,
+      "loss": 0.1976,
+      "step": 1000
+    },
+    {
+      "epoch": 0.64,
+      "eval_loss": 0.1308828443288803,
+      "eval_runtime": 385.5506,
+      "eval_samples_per_second": 14.102,
+      "eval_steps_per_second": 0.882,
+      "eval_wer": 0.10391520487905322,
+      "step": 1000
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 4.422260239264865e-06,
+      "loss": 0.1845,
+      "step": 1100
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 4.062141327272612e-06,
+      "loss": 0.1885,
+      "step": 1200
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 3.7020224152803586e-06,
+      "loss": 0.1799,
+      "step": 1300
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 3.3419035032881062e-06,
+      "loss": 0.1749,
+      "step": 1400
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 2.981784591295853e-06,
+      "loss": 0.1887,
+      "step": 1500
+    },
+    {
+      "epoch": 0.97,
+      "eval_loss": 0.13062627613544464,
+      "eval_runtime": 383.0748,
+      "eval_samples_per_second": 14.193,
+      "eval_steps_per_second": 0.888,
+      "eval_wer": 0.104032866934021,
+      "step": 1500
+    },
+    {
+      "epoch": 1.03,
+      "learning_rate": 2.6216656793036003e-06,
+      "loss": 0.1917,
+      "step": 1600
+    },
+    {
+      "epoch": 1.1,
+      "learning_rate": 2.261546767311348e-06,
+      "loss": 0.1737,
+      "step": 1700
+    },
+    {
+      "epoch": 1.16,
+      "learning_rate": 1.9014278553190947e-06,
+      "loss": 0.1758,
+      "step": 1800
+    },
+    {
+      "epoch": 1.22,
+      "learning_rate": 1.541308943326842e-06,
+      "loss": 0.174,
+      "step": 1900
+    },
+    {
+      "epoch": 1.29,
+      "learning_rate": 1.181190031334589e-06,
+      "loss": 0.18,
+      "step": 2000
+    },
+    {
+      "epoch": 1.29,
+      "eval_loss": 0.13111014664173126,
+      "eval_runtime": 382.6746,
+      "eval_samples_per_second": 14.208,
+      "eval_steps_per_second": 0.888,
+      "eval_wer": 0.10379754282408542,
+      "step": 2000
+    },
+    {
+      "epoch": 1.35,
+      "learning_rate": 8.210711193423363e-07,
+      "loss": 0.1667,
+      "step": 2100
+    },
+    {
+      "epoch": 1.42,
+      "learning_rate": 4.609522073500836e-07,
+      "loss": 0.1774,
+      "step": 2200
+    },
+    {
+      "epoch": 1.48,
+      "learning_rate": 1.0083329535783077e-07,
+      "loss": 0.1779,
+      "step": 2300
+    },
+    {
+      "epoch": 1.5,
+      "step": 2327,
+      "total_flos": 1.6280526756920635e+19,
+      "train_loss": 0.19505117766376623,
+      "train_runtime": 7387.6715,
+      "train_samples_per_second": 10.08,
+      "train_steps_per_second": 0.315
+    },
+    {
+      "epoch": 1.5,
+      "step": 2327,
+      "total_flos": 1.6280526756920635e+19,
+      "train_loss": 0.0,
+      "train_runtime": 12.046,
+      "train_samples_per_second": 412.128,
+      "train_steps_per_second": 6.475
+    },
+    {
+      "epoch": 1.5,
+      "step": 2327,
+      "total_flos": 1.6280526756920635e+19,
+      "train_loss": 0.0,
+      "train_runtime": 5.9011,
+      "train_samples_per_second": 841.279,
+      "train_steps_per_second": 26.436
+    }
+  ],
+  "max_steps": 156,
+  "num_train_epochs": 1,
+  "total_flos": 1.6280526756920635e+19,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cba0407afeb94e5c61fb60760555d3e67958b71b4ae430dc28c3be8ab33d4a2d
+size 3055

vocab.json

@@ -0,0 +1 @@
+{"a": 1, "b": 2, "c": 3, "d": 4, "e": 5, "f": 6, "g": 7, "h": 8, "i": 9, "j": 10, "k": 11, "l": 12, "m": 13, "n": 14, "o": 15, "p": 16, "q": 17, "r": 18, "s": 19, "t": 20, "u": 21, "v": 22, "w": 23, "x": 24, "y": 25, "z": 26, "å": 27, "æ": 28, "ø": 29, "|": 0, "[UNK]": 30, "[PAD]": 31}