README.md

metadata

library_name: transformers
language:
  - hu
license: apache-2.0
base_model: facebook/wav2vec2-large-xlsr-53
tags:
  - automatic-speech-recognition
  - mozilla-foundation/common_voice_17_0
  - generated_from_trainer
datasets:
  - common_voice_17_0
metrics:
  - wer
model-index:
  - name: wav2vec2-large-xlsr-53-hungarian
    results:
      - task:
          name: Automatic Speech Recognition
          type: automatic-speech-recognition
        dataset:
          name: MOZILLA-FOUNDATION/COMMON_VOICE_17_0 - HU
          type: common_voice_17_0
          config: hu
          split: test
          args: 'Config: hu, Training split: train+validation, Eval split: test'
        metrics:
          - name: Wer
            type: wer
            value: 0.1727824914378453

wav2vec2-large-xlsr-53-hungarian

This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the MOZILLA-FOUNDATION/COMMON_VOICE_17_0 - HU dataset. It achieves the following results on the evaluation set:

• Loss: 0.1748
• Wer: 0.2997

The training and measured wer values ​​differ due to ignored characters.

Model Comparison with the previous best wav2vec model (eval on CV17)

Model name	WER	CER
jonatasgrosman/wav2vec2-large-xlsr-53-hungarian	46.199835320230555	9.85170677112479
sarpba/wav2vec2-large-xlsr-53-hungarian	17.27824914378453	3.151354554132789

Igonore characters on eval:

CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
                   "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
                   "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
                   "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽",
                   "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]

Intended uses & limitations

More information needed

Train & Evaluation

Trained with transformers example pytorch script

Eval:

import torch
import librosa
import re
import warnings
from datasets import load_dataset
import evaluate
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

LANG_ID = "hu"
MODEL_ID = "jonatasgrosman/wav2vec2-large-xlsr-53-hungarian"
DEVICE = "cuda"

CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
                   "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
                   "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
                   "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽",
                   "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]

test_dataset = load_dataset("mozilla-foundation/common_voice_17_0", LANG_ID, split="test")

wer = evaluate.load("wer")  # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/wer.py
cer = evaluate.load("cer")  # https://github.com/jonatasgrosman/wav2vec2-sprint/blob/main/cer.py


chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"

processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)
model = Wav2Vec2ForCTC.from_pretrained(MODEL_ID)
model.to(DEVICE)

# Preprocessing the datasets.
# We need to read the audio files as arrays
def speech_file_to_array_fn(batch):
    with warnings.catch_warnings():
        warnings.simplefilter("ignore")
        speech_array, sampling_rate = librosa.load(batch["path"], sr=16_000)
    batch["speech"] = speech_array
    batch["sentence"] = re.sub(chars_to_ignore_regex, "", batch["sentence"]).upper()
    return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

# Preprocessing the datasets.
# We need to read the audio files as arrays
def evaluate(batch):
    inputs = processor(batch["speech"], sampling_rate=16_000, return_tensors="pt", padding=True)

    with torch.no_grad():
        logits = model(inputs.input_values.to(DEVICE), attention_mask=inputs.attention_mask.to(DEVICE)).logits

    pred_ids = torch.argmax(logits, dim=-1)
    batch["pred_strings"] = processor.batch_decode(pred_ids)
    return batch

result = test_dataset.map(evaluate, batched=True, batch_size=8)

predictions = [x.upper() for x in result["pred_strings"]]
references = [x.upper() for x in result["sentence"]]

print(f"WER: {wer.compute(predictions=predictions, references=references) * 100}")
print(f"CER: {cer.compute(predictions=predictions, references=references) * 100}")

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 15.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.7968	1.0	758	0.2848	0.5295
0.2547	2.0	1516	0.1908	0.4222
0.1929	3.0	2274	0.1753	0.4000
0.1532	4.0	3032	0.1558	0.3710
0.1297	5.0	3790	0.1512	0.3536
0.1167	6.0	4548	0.1574	0.3514
0.101	7.0	5306	0.1483	0.3374
0.0859	8.0	6064	0.1490	0.3299
0.0791	9.0	6822	0.1523	0.3250
0.0702	10.0	7580	0.1608	0.3192
0.0629	11.0	8338	0.1664	0.3146
0.0559	12.0	9096	0.1641	0.3103
0.0527	13.0	9854	0.1665	0.3063
0.0468	14.0	10612	0.1691	0.3011
0.0443	15.0	11370	0.1748	0.2998

Framework versions

• Transformers 4.50.0.dev0
• Pytorch 2.6.0+cu124
• Datasets 3.3.2
• Tokenizers 0.21.0