Stt Kr Conformer Ctc Medium

🚀 Conformer-ctc-medium-ko

This model is a fine - tuned version of RIVA Conformer ASR Korean on the AI hub dataset. Conformer - based models, unlike attention - based models like Whisper, have the advantages of maintaining high performance during streaming and fast speed.

🚀 Quick Start

This Conformer - based model is fine - tuned from RIVA Conformer ASR Korean on the AI hub dataset. Unlike attention - based models such as Whisper, Conformer - based models do not experience a significant performance drop during streaming and are fast.

It was found that the Real - Time Factor (RTF) is about 0.05 on a V100 GPU and 0.35 on a 7 - core CPU. In the streaming test with an audio chunk size of 2 seconds, there is about a 20% performance degradation compared to using the entire audio, but the performance is still acceptable.

Additionally, in domains such as customer service voice (non - open domain), adding KenLM significantly improved the Word Error Rate (WER) from 13.45 to 5.27. However, in other domains, adding KenLM did not lead to a significant performance improvement.

The code for streaming and the code including the denoise model can be found on the following GitHub repository: https://github.com/SUNGBEOMCHOI/Korean-Streaming-ASR

✨ Features

• Streaming - friendly: Maintains relatively stable performance during streaming compared to attention - based models.
• Fast speed: Achieves low RTF values on both GPU and CPU.
• Domain - specific performance improvement: Adding KenLM can significantly improve WER in specific domains.

📚 Documentation

Training results

Dataset

🔧 Technical Details

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e - 05
• train_batch_size: 16
• eval_batch_size: 16
• num_train_epoch: 1
• sample_rate: 16000
• max_duration: 20.0

📄 License

This project is licensed under the Apache - 2.0 license.