A naukri.com initiative
Analyticsindiamag
1M
144
Image Credit: Analyticsindiamag
Hey Siri, You There?
- Apple has developed a technique for improving Automated Speech Recognition (ASR) on older devices while reducing computational time.
- ASR systems must identify rare and user-specific words, which the new system tackles through neural context biasing (NCB).
- NCB processing consumes significant computational and memory resources, but Apple's model uses vector quantisation to overcome this.
- Results of the new approach show 20% less processing time and a 71% reduction in errors observed.
- Earlier this year, Apple introduced the DenoisingLM (DLM) to achieve an accurate rate for ASR.
- However, voice assistants like Siri were far from perfect, even with Apple's AI system, especially with certain regions' accents.
- Despite concerns about quantisation loss, Apple's use of finite scaling quantisation (FSQ) to improve ASR is being seen as promising.
- While Apple Intelligence, inclusive of Siri, is ideal, it’s still less mature than competitors, such as Whisper, from the likes of OpenAI.
- Other techniques have been created by Moonshine to reduce computing demands for accurate ASRs.
- Assembly AI’s latest ASR model, Universal-2, also improves proper noun recognition and formatting accuracy.
Read Full Article
8 Likes
For uninterrupted reading, download the app