wav2vec2 large xlsr kazakh

Introduction

The WAV2VEC2-LARGE-XLSR-53-KAZAKH model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 for Automatic Speech Recognition (ASR) in the Kazakh language. It was trained using the Kazakh Speech Corpus v1.1.

Architecture

The model is based on the Wav2Vec2 architecture, specifically the facebook/wav2vec2-large-xlsr-53 variant, which is designed for multilingual and cross-lingual speech recognition tasks. The model processes audio inputs sampled at 16kHz.

Training

The model was trained using the Kazakh Speech Corpus v1.1. The fine-tuning process involved adapting the pre-trained Wav2Vec2 model to recognize and transcribe Kazakh speech, achieving a Word Error Rate (WER) of 19.65%.

Guide: Running Locally

1. Setup Environment: Ensure you have Python, PyTorch, Torchaudio, and Hugging Face Transformers installed.

2. Load Model and Processor:

   from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
   processor = Wav2Vec2Processor.from_pretrained("wav2vec2-large-xlsr-kazakh")
   model = Wav2Vec2ForCTC.from_pretrained("wav2vec2-large-xlsr-kazakh")
   

3. Prepare Data:

   • Use the Kazakh Speech Corpus v1.1.
   • Ensure audio files are resampled to 16kHz.

4. Run Inference:

   inputs = processor(audio_input, sampling_rate=16000, return_tensors="pt", padding=True)
   with torch.no_grad():
       logits = model(inputs.input_values).logits
   predicted_ids = torch.argmax(logits, dim=-1)
   transcription = processor.batch_decode(predicted_ids)
   

5. Evaluation: Use the wer metric from Hugging Face's datasets library to evaluate the model on a test set.

For faster processing, consider using cloud GPUs such as those available on AWS, GCP, or Azure.

License

The model is licensed under the Apache 2.0 License.