lightshifted/whisper_pretraining_sw_yo

Fine-tuned Whisper models to build state-of-the art speech recognition systems in the Swahili and Yoruba languages. # Fine-Tune Whisper for Speech Recognition In this project, we fine-tune the Whisper model and develop cutting-edge speech recognition systems for the Swahili language. The techniques we used can also be applied to other languages. Further, we evaluate our final model on the "test" split of the Common Voice 11 dataset for Swahili, allowing us to assess the effectiveness of our approach and compare it with other models. ## Overview We fine-tuned the Whisper model to improve the accuracy of speech recognition in the Swahili language. To do this, we used a large dataset of audio samples in this languages to train the model and optimize its performance. We also developed a user-friendly interface for the speech recognition system that allows users to easily interact with the model and obtain accurate transcriptions of their speech in real-time. This interface can be easily integrated into existing speech recognition applications and platforms to expand their capabilities and make them more accessible to users of the Swahili and Yoruba languages. The interface is accessible here. Our long term objective is to create a community of users and developers who can contribute to the project by providing feedback, suggestions, and additional data to further improve the performance of the speech recognition system. This community allows us to continuously improve the model and make it more effective for users of the Swahili language. ## Model Weights The pretrained model weights from this project are stored in HuggingFace repositories: | Model | Repository | |--------|------------------| | whisper-small-sw | link | | whisper-medium-sw | link | | whisper-large-v2-sw | link | ## Set Up an Environment ## Running the Script 1. **Clone model repository** The steps for running training with a Python script assume that you are SSH'd into your GPU device and have set up your environment according to the previous section Set Up an Environment. We chose to host our pretr …