OpenAI Releases Open Source Whisper Speech Recognition Model
OpenAI introduces Whisper, a highly robust open-source neural network for speech recognition that approaches human-level accuracy. The model handles various accents and languages but requires significant processing power to run locally.
OpenAI introduces Whisper, an open-source neural network that brings human-level accuracy and robustness to English speech recognition. Unlike closed-source virtual assistants or phone-based tools, this solution gives developers full access to the technology for major projects without relying on proprietary ecosystems.
The system handles fast speech and heavy accents with remarkable ease, drawing on a massive training dataset of 680,000 hours of multilingual audio. It processes audio by splitting it into 30-second chunks, converting them into spectrograms, and using encoders and decoders to predict the text accurately.
While the generalized training causes it to slightly underperform on specific benchmarks compared to specialized models, it excels at handling random, real-world speech. Developers should note that hardware requirements are steep, with model sizes ranging from 39 megabytes for the "tiny" version to over 1.5 gigabytes for the "large" variant, keeping it off microcontrollers like the Arduino for now.