A team just made OpenAI Whisper 6x faster, 49% smaller, while keeping 99% of the accuracy.
As the size of pre-trained speech recognition models increases, running these large models in low-latency or resource-constrained environments becomes challenging. In this work, we leverage pseudo-labelling to assemble a large-scale open-source dataset which we use to distill the Whisper model into a smaller variant, called Distil-Whisper. Using a simple word error rate (WER) heuristic, we select only the highest quality pseudo-labels for training. The distilled model is 5.8 times faster with 51% fewer parameters, while performing to within 1% WER on out-of-distribution test data in a zero-shot transfer setting. Distil-Whisper maintains the robustness of the Whisper model to difficult acoustic conditions, while being less prone to hallucination errors on long-form audio. Distil-Whisper is designed to be paired with Whisper for speculative decoding, yielding a 2 times speed-up while mathematically ensuring the same outputs as the original model. To facilitate further research in this domain, we make our training code, inference code and models publicly accessible.
CopyAI – Build and automate your team’s workflows
Workflows by Copy.ai are a brand new solution for teams who are hungry for results at scale. Simply type the workflow you want, verify the results, and run it in bulk.
🌶️🚨How to (almost) protect the instructions of your CustomGPT?
With the advent of a GPTs store, developers are concerned that their unique instructions, which add value to GPTs, might be copied or leaked during interactions with users. This led to the development of various protection methods, though four versions have already been cracked. The latest solution, version 5.0, emphasizes language-based protection rather than code, aligning with the idea that language is the new code.
AI outperforms conventional weather forecasting methods for the first time
The article from Financial Times discusses a significant breakthrough in weather forecasting achieved by Google DeepMind’s AI model, GraphCast. For the first time, this AI system has outperformed traditional forecasting methods. The evaluation showed that GraphCast was more accurate than the European Centre for Medium-range Weather Forecasts (ECMWF), the leading conventional system, in predicting weather up to 10 days ahead, excelling in 90% of the 1,380 metrics used.
GraphCast employs a machine-learning architecture known as graph neural network and was trained on over 40 years of ECMWF data. It produces 10-day forecasts in just a minute on a single Google TPU v4 cloud computer, which is far more energy-efficient compared to the energy-intensive process of conventional methods. An example of its accuracy was highlighted with Hurricane Lee, where GraphCast predicted its landfall in Nova Scotia nine days in advance, three days earlier than traditional methods.
Despite these advancements, AI models like GraphCast still face challenges, as shown in the prediction of Hurricane Otis’s intensification. Future developments include combining AI with traditional methods to improve accuracy and account for climate change. The UK Met Office, in collaboration with the Alan Turing Institute, is also working on developing its own graph neural network for weather forecasting.
Prompt of the Day 🍭: One-Click Book Writer
Click on the image to load the full prompt in ChatGPT (requires Superpower ChatGPT)
Hope you enjoyed today’s newsletter
⚡️ Join over 200,000 people using the Superpower ChatGPT extension on Chrome and Firefox.