Naver Applies 'HyperCLOVA' Technology to Voice Recognition AI

[Asia Economy Reporter Kang Nahum] Naver announced on the 8th that it has significantly enhanced the performance of its speech recognition AI engine by applying the deep learning technology of 'HyperCLOVA'.

Conventional AI deep learning has been conducted through a 'supervised learning' method, which trains pairs of data and their descriptive 'labels'. For example, an image of a cat is trained together with the label 'cat'.

To develop speech recognition AI, both audio and the text within the audio must be trained together. This 'data labeling' is done manually by humans, requiring significant time and cost to build the training dataset.

On the other hand, 'self-supervised learning' is a cutting-edge deep learning technique that can learn from data itself without labels, greatly improving training efficiency. This is also the core technology of Naver's recently unveiled hyperscale AI, 'HyperCLOVA'.

Naver explained that it applied the self-supervised learning method to its AI speech recognition engine 'NEST (Neural End-to-end Speech Transcriber)', increasing speech recognition accuracy by about 30% compared to before. With the new training method, transcription work to verify text in audio data is minimized while enabling more accurate AI training than before, drastically reducing the model's training time and cost.

The NEST engine is optimized to recognize complex long-form speech expressions rather than short speech commands, and was first unveiled by Naver in April last year.

The upgraded NEST engine was first integrated into 'CLOVA Note'. CLOVA Note, a voice recording service launched by Naver at the end of last year, has received positive feedback from users based on its high recognition rate.

Naver plans to install this engine in various services utilizing speech recognition AI, such as CLOVA Care Call, to enhance user convenience.

This content was produced with the assistance of AI translation services.