--- library_name: transformers language: - ru license: apache-2.0 base_model: jonatasgrosman/wav2vec2-large-xlsr-53-russian tags: - hf-asr-leaderboard - generated_from_trainer datasets: - mozilla-foundation/common_voice_17_0 metrics: - wer model-index: - name: Wav2vec2-large ru - slowlydoor results: - task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: Common Voice 17.0 type: mozilla-foundation/common_voice_17_0 config: ru split: None args: 'config: ru, split: test' metrics: - name: Wer type: wer value: 22.3988525667842 --- # Wav2vec2-large ru - slowlydoor ([Automatic Speech Recognition](https://github.com/SlowlyDoor/Automatic-Speech-Recognition)) This model is a fine-tuned version of [jonatasgrosman/wav2vec2-large-xlsr-53-russian](https://huggingface.co/jonatasgrosman/wav2vec2-large-xlsr-53-russian) on the Common Voice 17.0 dataset. It achieves the following results on the evaluation set: - Loss: 0.2124 - Wer: 22.3989 - Cer: 4.8036 - Ser: 75.4264 ## Model description More information needed ## Intended uses & limitations More information needed ## Training and evaluation data More information needed ## Training procedure ### Training hyperparameters The following hyperparameters were used during training: - learning_rate: 0.0001 - train_batch_size: 8 - eval_batch_size: 4 - seed: 42 - optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments - lr_scheduler_type: linear - num_epochs: 5 - mixed_precision_training: Native AMP ### Training code ```bash pip install datasets librosa scikit-learn torch torchaudio evaluate jiwer nltk pip install --upgrade datasets ``` ```python from huggingface_hub import login from datasets import load_dataset, DatasetDict from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2CTCTokenizer, Wav2Vec2Processor, Wav2Vec2ForCTC, TrainingArguments, Trainer from datasets import load_dataset, Audio import torch import torchaudio import re import evaluate import numpy as np from dataclasses import dataclass, field from typing import Any, Dict, List, Optional, Union login("***") common_voice = DatasetDict() common_voice["train"] = load_dataset("mozilla-foundation/common_voice_17_0", "ru", split="train") common_voice["test"] = load_dataset("mozilla-foundation/common_voice_17_0", "ru", split="test") common_voice = common_voice.remove_columns(["accent", "age", "client_id", "down_votes", "gender", "locale", "path", "segment", "up_votes"]) common_voice = common_voice.cast_column("audio", Audio(sampling_rate=16000)) tokenizer = Wav2Vec2CTCTokenizer.from_pretrained("jonatasgrosman/wav2vec2-large-xlsr-53-russian") feature_extractor = Wav2Vec2FeatureExtractor(feature_size=1, sampling_rate=16000, padding_value=0.0, return_attention_mask=True) processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer) def prepare_dataset(batch): audio = batch["audio"] # batched output is "un-batched" batch["input_values"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0] batch["input_length"] = len(batch["input_values"]) with processor.as_target_processor(): batch["labels"] = processor(batch["sentence"]).input_ids return batch common_voice = common_voice.map(prepare_dataset, remove_columns=common_voice["train"].column_names, num_proc=2) wer_metric = evaluate.load("wer") cer_metric = evaluate.load("cer") def compute_metrics(pred): pred_logits = pred.predictions pred_ids = np.argmax(pred_logits, axis=-1) pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id pred_str = processor.batch_decode(pred_ids, skip_special_tokens=True) label_str = processor.batch_decode(pred.label_ids, group_tokens=False, skip_special_tokens=True) pairs = [(ref.strip(), hyp.strip()) for ref, hyp in zip(label_str, pred_str)] pairs = [(ref, hyp) for ref, hyp in pairs if len(ref) > 0] if len(pairs) == 0: return {"wer": 1.0, "cer": 1.0, "ser": 1.0} label_str, pred_str = zip(*pairs) wer = 100 * wer_metric.compute(predictions=pred_str, references=label_str) cer = 100 * cer_metric.compute(predictions=pred_str, references=label_str) incorrect_sentences = sum([ref != pred for ref, pred in zip(label_str, pred_str)]) ser = 100 * incorrect_sentences / len(label_str) return { "wer": wer, "cer": cer, "ser": ser } model = Wav2Vec2ForCTC.from_pretrained( "jonatasgrosman/wav2vec2-large-xlsr-53-russian", ctc_loss_reduction="mean", pad_token_id=processor.tokenizer.pad_token_id, ) @dataclass class DataCollatorCTCWithPadding: processor: Wav2Vec2Processor padding: Union[bool, str] = True max_length: Optional[int] = None max_length_labels: Optional[int] = None 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 lengths 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, max_length=self.max_length, 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, max_length=self.max_length_labels, 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 data_collator = DataCollatorCTCWithPadding(processor=processor, padding=True) training_args = TrainingArguments( output_dir="/content/drive/MyDrive/models/wav2vec2-large-ru-5ep", logging_dir="/content/drive/MyDrive/models/wav2vec2-large-ru-5ep", group_by_length=True, per_device_train_batch_size=8, per_device_eval_batch_size=4, eval_strategy="steps", logging_strategy="steps", save_strategy="steps", num_train_epochs=5, logging_steps=25, eval_steps=500, save_steps=500, fp16=True, optim="adamw_torch_fused", torch_compile=True, gradient_checkpointing=True, learning_rate=1e-4, weight_decay=0.005, report_to=["tensorboard"], push_to_hub=False ) trainer = Trainer( model=model, data_collator=data_collator, args=training_args, compute_metrics=compute_metrics, train_dataset=common_voice["train"], eval_dataset=common_voice["test"], tokenizer=processor, ) trainer.train() ``` ### Training results | Training Loss | Epoch | Step | Validation Loss | Wer | Cer | Ser | |:-------------:|:------:|:-----:|:---------------:|:-------:|:------:|:-------:| | 0.3421 | 0.1516 | 500 | 0.2593 | 27.7416 | 6.2518 | 81.6311 | | 0.2979 | 0.3032 | 1000 | 0.2741 | 27.9854 | 6.3745 | 82.2290 | | 0.2787 | 0.4548 | 1500 | 0.2538 | 27.3041 | 6.0743 | 81.1998 | | 0.325 | 0.6064 | 2000 | 0.2701 | 29.4006 | 6.5501 | 83.6503 | | 0.3048 | 0.7580 | 2500 | 0.2435 | 27.0914 | 6.0148 | 80.8077 | | 0.294 | 0.9096 | 3000 | 0.2495 | 26.9503 | 5.9946 | 80.9939 | | 0.2648 | 1.0612 | 3500 | 0.2675 | 26.8356 | 6.0261 | 80.8175 | | 0.2691 | 1.2129 | 4000 | 0.2372 | 26.1220 | 5.8259 | 80.2294 | | 0.2245 | 1.3645 | 4500 | 0.2394 | 26.1603 | 5.8315 | 80.3470 | | 0.2738 | 1.5161 | 5000 | 0.2388 | 26.0420 | 5.7826 | 79.9941 | | 0.2767 | 1.6677 | 5500 | 0.2330 | 25.8089 | 5.7248 | 79.5138 | | 0.2689 | 1.8193 | 6000 | 0.2284 | 25.7312 | 5.6832 | 79.6216 | | 0.2571 | 1.9709 | 6500 | 0.2370 | 25.3403 | 5.6065 | 79.3080 | | 0.2479 | 2.1225 | 7000 | 0.2372 | 25.2065 | 5.5776 | 78.9943 | | 0.2021 | 2.2741 | 7500 | 0.2284 | 24.8718 | 5.4638 | 78.6610 | | 0.1864 | 2.4257 | 8000 | 0.2280 | 24.8132 | 5.4340 | 78.8669 | | 0.1953 | 2.5773 | 8500 | 0.2237 | 24.4941 | 5.3856 | 78.3670 | | 0.195 | 2.7289 | 9000 | 0.2190 | 24.2658 | 5.2770 | 77.8279 | | 0.1829 | 2.8805 | 9500 | 0.2194 | 24.2443 | 5.2697 | 77.8671 | | 0.1457 | 3.0321 | 10000 | 0.2205 | 24.2587 | 5.2398 | 77.8279 | | 0.1435 | 3.1837 | 10500 | 0.2223 | 23.7985 | 5.1608 | 77.1613 | | 0.1435 | 3.3354 | 11000 | 0.2219 | 23.6551 | 5.1230 | 76.9065 | | 0.1752 | 3.4870 | 11500 | 0.2186 | 23.4829 | 5.0767 | 76.5438 | | 0.1793 | 3.6386 | 12000 | 0.2232 | 23.4339 | 5.0977 | 76.4556 | | 0.1682 | 3.7902 | 12500 | 0.2133 | 23.1853 | 5.0090 | 76.0929 | | 0.1607 | 3.9418 | 13000 | 0.2135 | 22.7610 | 4.9091 | 75.7597 | | 0.1463 | 4.0934 | 13500 | 0.2138 | 22.8495 | 4.9314 | 76.1125 | | 0.1654 | 4.2450 | 14000 | 0.2138 | 22.6379 | 4.8814 | 75.7008 | | 0.1586 | 4.3966 | 14500 | 0.2173 | 22.6678 | 4.8705 | 75.5342 | | 0.1438 | 4.5482 | 15000 | 0.2166 | 22.5411 | 4.8437 | 75.5342 | | 0.1645 | 4.6998 | 15500 | 0.2146 | 22.4658 | 4.8308 | 75.3774 | | 0.1254 | 4.8514 | 16000 | 0.2124 | 22.3989 | 4.8036 | 75.4264 | ### Framework versions - Transformers 4.52.2 - Pytorch 2.6.0+cu124 - Datasets 3.6.0 - Tokenizers 0.21.1