add create lm scripts

Lemswasabi_tuudle_rtl-14h_test_eval_results.txt

@@ -0,0 +1,2 @@
+WER: 0.09642258244829514
+CER: 0.02028596961572833

add_eos_token.py

@@ -0,0 +1,18 @@
+#!/usr/bin/env python3
+#
+#  Created by lemswasabi on 24/05/2022.
+#  Copyright © 2022 letzspeak. All rights reserved.
+#
+
+with open("5gram.arpa", "r") as read_file, open("5gram_correct.arpa", "w") as write_file:
+  has_added_eos = False
+  for line in read_file:
+    if not has_added_eos and "ngram 1=" in line:
+      count=line.strip().split("=")[-1]
+      write_file.write(line.replace(f"{count}", f"{int(count)+1}"))
+    elif not has_added_eos and "<s>" in line:
+      write_file.write(line)
+      write_file.write(line.replace("<s>", "</s>"))
+      has_added_eos = True
+    else:
+      write_file.write(line)

create_lm_decoder.py

@@ -0,0 +1,26 @@
+#!/usr/bin/env python3
+#
+#  Created by lemswasabi on 24/05/2022.
+#  Copyright © 2022 letzspek. All rights reserved.
+#
+
+from transformers import AutoProcessor
+from transformers import Wav2Vec2ProcessorWithLM
+from pyctcdecode import build_ctcdecoder
+
+processor = AutoProcessor.from_pretrained("./")
+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="5gram_correct.arpa",
+)
+
+processor_with_lm = Wav2Vec2ProcessorWithLM(
+    feature_extractor=processor.feature_extractor,
+    tokenizer=processor.tokenizer,
+    decoder=decoder
+)
+
+processor_with_lm.save_pretrained("./")

create_text_corpus.py

@@ -0,0 +1,35 @@
+#!/usr/bin/env python3
+#
+#  Created by lemswasabi on 24/05/2022.
+#  Copyright © 2022 letzspeak. All rights reserved.
+#
+
+import glob
+import re
+import textract
+
+
+chars_to_ignore_regex = '[,?.!;:"“%‘„”�—’…–]'
+
+def replace_chars(text, char, replace_char):
+    return re.sub(char, replace_char, text.lower())
+
+def ignore_chars(sentence):
+    return re.sub(chars_to_ignore_regex, "", text.lower())
+
+corpus = []
+
+for text_file in glob.glob("/home/lemswasabi/corpus/chamber_text_corpus/**/*.doc", recursive=True):
+    try:
+        text = textract.process(text_file).decode("utf-8")
+        text = replace_chars(text, "’", "'")
+        text = replace_chars(text, "‘", "'")
+        text = replace_chars(text, "-", " ")
+        text = replace_chars(text, "\\n", " ")
+        text = ignore_chars(text)
+        corpus.append(text.strip())
+    except textract.exceptions.ShellError:
+        continue
+
+with open("chamber_text.txt", "w") as f:
+    f.write(" ".join(corpus))

eval.py

@@ -0,0 +1,137 @@
+#!/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}\nCER: {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())
+
+    # 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["sentence"])
+        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)