OpenAI has launched its latest Whisper model, the Whisper V3 Turbo, which significantly enhances transcription capabilities. This newly released model offers transcription speeds that are eight times faster than its predecessor, large-v3 while maintaining a comparable level of accuracy.
OpenAI has released new Whisper models! 👀 Yesterday, @OpenAI updated their Github and added a new Whisper V3 Turbo model!
— Philipp Schmid (@_philschmid) October 1, 2024
The turbo model is an optimized version of large-v3 that offers 8x faster transcription speed with minimal degradation in accuracy (no benchmarks yet) with… pic.twitter.com/ixSFdLgTO3
The Whisper V3 Turbo is not only faster but also more efficient, being roughly half the size of the previous version. This optimisation allows for easier deployment across various platforms, making high-speed transcription accessible to a broader audience. Although no official benchmarks have been released yet to quantify the accuracy differences, early indications suggest minimal degradation in performance.
The announcement was made via OpenAI’s official GitHub repository, where the company continues to expand its suite of AI tools. The Whisper models, known for their versatility in handling different languages and accents, are widely used in applications ranging from automated customer service to content creation.
Developed by OpenAI, Whisper boasts an impressive ability to convert spoken language into written text across over 99 languages, making it one of the most versatile ASR systems available today. Its robustness in handling various accents, background noise, and technical language sets it apart from many other speech recognition tools.
At its core, Whisper utilises an encoder-decoder Transformer architecture, trained on a massive dataset of 680,000 hours of multilingual and multitask supervised data. This extensive training allows Whisper to achieve near-human level accuracy in English speech recognition and outperform many specialized models in zero-shot performance across diverse datasets.
The system processes audio in 30-second chunks, converting them into log-Mel spectrograms, and employs special tokens to perform tasks such as language identification, phrase-level timestamps, and translation.
One of Whisper’s key strengths lies in its open-source availability, with models and inference code accessible to developers and researchers worldwide. This openness has fostered a community of innovation, allowing for customisation and improvement of the technology. Whisper is also available through OpenAI’s API, making it easy for developers to integrate into their applications and services.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.