brianyan918_iwslt22_dialect_train_asr_conformer_ctc0.3_lr2e 3_warmup15k_newspecaug

Introduction

The ESPnet2 ASR Model is an automatic speech recognition (ASR) system developed by Brian Yan using the iwslt22_dialect recipe in ESPnet. It is designed for speech processing tasks, leveraging a Conformer architecture with CTC (Connectionist Temporal Classification) for improved performance.

Architecture

The model utilizes a Conformer architecture, which combines convolutional neural networks (CNNs) and transformers to capture both local and global dependencies in audio data. Key components include:

• Encoder: 12 blocks with 256 output size, using 4 attention heads, convolutional layers, and self-attention mechanisms.
• Decoder: 6 blocks with 2048 linear units and similar attention mechanisms.
• CTC Weight: 0.3, used to balance the CTC loss with the main sequence-to-sequence loss.

Training

Training was conducted using the iwslt22_dialect dataset with a focus on optimizing parameters such as learning rate (0.002) and utilizing a warm-up learning rate scheduler with 15000 warm-up steps. Data augmentation techniques like SpecAugment were applied to improve generalization.

Guide: Running Locally

To run the model locally, follow these steps:

1. Environment Setup:

   • Clone the ESPnet repository:

     git clone https://github.com/espnet/espnet.git
     cd espnet
     

   • Check out the specific commit:

     git checkout 77fce65312877a132bbae01917ad26b74f6e2e14
     

   • Install dependencies:

     pip install -e .
     

2. Run the ASR Recipe:

   • Navigate to the example directory:

     cd egs2/iwslt22_dialect/asr1
     

   • Execute the run script:

     ./run.sh --skip_data_prep false --skip_train true --download_model espnet/brianyan918_iwslt22_dialect_train_asr_conformer_ctc0.3_lr2e-3_warmup15k_newspecaug
     

3. Optional: Use cloud GPUs from providers like AWS, GCP, or Azure for faster processing.

License

The model is released under the CC-BY-4.0 license, allowing for sharing and adaptation with appropriate credit.