Wav2vec2 Large Xlsr 53 Turkish

🚀 Wav2Vec2-Large-XLSR-53-Turkish

This model is fine-tuned from facebook/wav2vec2-large-xlsr-53 on Turkish using the Common Voice. It's designed for automatic speech recognition in Turkish.

Model Information

Results

🚀 Quick Start

When using this model, make sure that your speech input is sampled at 16kHz.

✨ Features

• Fine-tuned on Turkish using the Common Voice dataset.
• Can be used directly for speech recognition without a language model.

📦 Installation

No specific installation steps are provided in the original document, so this section is skipped.

💻 Usage Examples

Basic Usage

The model can be used directly (without a language model) as follows:

import torch
import torchaudio
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
from unicode_tr import unicode_tr

test_dataset = load_dataset("common_voice", "tr", split="test[:2%]")

processor = Wav2Vec2Processor.from_pretrained("aniltrkkn/wav2vec2-large-xlsr-53-turkish")
model = Wav2Vec2ForCTC.from_pretrained("aniltrkkn/wav2vec2-large-xlsr-53-turkish")

resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
	speech_array, sampling_rate = torchaudio.load(batch["path"])
	batch["speech"] = resampler(speech_array).squeeze().numpy()
	return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)
inputs = processor(test_dataset["speech"][:2], sampling_rate=16_000, return_tensors="pt", padding=True)

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

predicted_ids = torch.argmax(logits, dim=-1)

print("Prediction:", processor.batch_decode(predicted_ids))
print("Reference:", test_dataset["sentence"][:2])

Advanced Usage

The model can be evaluated as follows on the Turkish test data of Common Voice.

import torch
import torchaudio
from datasets import load_dataset, load_metric
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
import re

test_dataset = load_dataset("common_voice", "tr", split="test")
wer = load_metric("wer")

processor = Wav2Vec2Processor.from_pretrained("aniltrkkn/wav2vec2-large-xlsr-53-turkish")
model = Wav2Vec2ForCTC.from_pretrained("aniltrkkn/wav2vec2-large-xlsr-53-turkish")
model.to("cuda")

chars_to_ignore_regex = '[\\,\\?\\.\\!\\-\\;\\:\\\"\\“]'
resampler = torchaudio.transforms.Resample(48_000, 16_000)

# Preprocessing the datasets.
# We need to read the aduio files as arrays
def speech_file_to_array_fn(batch):
	batch["sentence"] = str(unicode_tr(re.sub(chars_to_ignore_regex, "", batch["sentence"])).lower())
	speech_array, sampling_rate = torchaudio.load(batch["path"])
	batch["speech"] = resampler(speech_array).squeeze().numpy()
	return batch

test_dataset = test_dataset.map(speech_file_to_array_fn)

# Preprocessing the datasets.
# We need to read the aduio 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("cuda"), attention_mask=inputs.attention_mask.to("cuda")).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)

print("WER: {:2f}".format(100 * wer.compute(predictions=result["pred_strings"], references=result["sentence"])))

Test Result: 17.46 %

🔧 Technical Details

• Case Conversion: The unicode_tr package is used for converting sentences to lower case since the regular lower() method does not work well with Turkish.
• Training Approach: Since training data is very limited for Turkish, all data is employed with a K-Fold (k=5) training approach. The best model out of the 5 trainings is uploaded.
• Training Arguments: --num_train_epochs="30" \ --per_device_train_batch_size="32" \ --evaluation_strategy="steps" \ --activation_dropout="0.055" \ --attention_dropout="0.094" \ --feat_proj_dropout="0.04" \ --hidden_dropout="0.047" \ --layerdrop="0.041" \ --learning_rate="2.34e-4" \ --mask_time_prob="0.082" \ --warmup_steps="250" \

All trainings took ~20 hours with a GeForce RTX 3090 Graphics Card.

📄 License

This model is licensed under the Apache-2.0 license.