The goal of this project is to transfer a speaker recognition technology from the Idiap research institute to KeyLemon SA. Currently, KeyLemon only owns a face recognition technology and is willing to complement its technology portfolio to be more competitive, to win new prospects and to develop new products based on speaker recognition or on
multi-modal face and speaker recognition. As a matter of fact, several companies (such as Credit Suisse) have expressed their interest into such technologies.
Recently, Idiap proposed a novel approach for automatic speaker recognition. It has been shown to be more noise-robust and computationally efficient than existing approaches. This makes it very suitable more particularly for applications intended for mobile devices, which are liable to be used in very noisy scenarios and possess comparatively limited computing power. However, the gain of processing power is also a significant advantage for desktop (or cloud-based) infrastructures, as it will allow allocating more computing power on subsequent tasks.
The proposed approach is based on a set of binary-valued and parts-based features, contrary to standard cepstral features used in speaker recognition, which are holistic and real-valued. The chief advantages and improvements of this new approach are as follows:
1. The features are very simple and easy to compute. They involve only comparison and addition operations. This contrasts with the standard cepstral features, which involve many more computations.
2. The features are considerably more robust to different types of noise that are encountered by speech-related applications in a real scenario (for example, in mobile phone applications), compared to standard cepstral features.
The main objectives of this NCCR Transfer Project are:
• for Idiap (1) to implement and to benchmark the proposed approach into a professional software library to facilitate technology transfer, (2) to develop a demonstration prototype,
• for KeyLemon (1) to develop a product prototype with the transferred speaker recognition technology, (2) to conduct a test field with a potential client, (3) to take a position on the licensing of Idiap speaker recognition technology.