Wav2vec2 Russian

🚀 wav2vec2-russian

This is a model for audio recognition. The results of this model can be further processed by my text correction network, UrukHan/t5-russian-spell.

🚀 Quick Start

You can download an audio file in WAV format for recognition. The result can be corrected by another network of mine, which corrects errors, adds punctuation, and fixes numbers. UrukHan/t5-russian-spell

Here is an example of running the network in Colab: Colab Example

from transformers import AutoModelForCTC, Wav2Vec2Processor
model = AutoModelForCTC.from_pretrained("UrukHan/wav2vec2-russian")
processor = Wav2Vec2Processor.from_pretrained("UrukHan/wav2vec2-russian")
def map_to_result(batch):
  with torch.no_grad():
    input_values = torch.tensor(batch["input_values"]).unsqueeze(0) #, device="cuda"
    logits = model(input_values).logits
  pred_ids = torch.argmax(logits, dim=-1)
  batch = processor.batch_decode(pred_ids)[0]
  return batch
 
map_to_result()

💻 Usage Examples

Basic Usage

The following table shows the comparison between the output of wav2vec2 and the output of the spell corrector:

🔧 Technical Details

You can understand the model training process, data processing, and dataset creation in Colab: Training Example