GPT-4: Whisper JAX is an optimized JAX implementation of OpenAI’s Whisper Model that runs over 70x faster than the PyTorch code. It is compatible with CPU, GPU, and TPU, and offers a quick-start guide for transcribing 30 minutes of audio in approximately 30 seconds. The model supports half-precision, batching, speech translation, and timestamp prediction. It can be used with Hugging Face Transformers and T5x partitioning for advanced parallelization techniques. Benchmarks show significant speed improvements over OpenAI and Transformers implementations, making it the fastest Whisper implementation available.
Read more at GitHub…