Speaker Verification

Trinity College Dublin Speaker Ageing Database

The Trinity College Dublin Speaker Ageing (TCDSA) Database was designed primarily to investigate the effect of ageing-related vocal change on speaker verification

The main portion of this database contains speech recordings of 26 adults (15 males and 11 females) across a time span of between 25-58 years per speaker.

Also included is a set of 120 development speakers, balanced across age group, gender and accent.

The TCDSA database is made available free of charge for academic use only.

Contact Finnian Kelly (kellyfp[a]tcd.ie) to obtain the data, or for further information.

Figure 1: A schematic of the TCDSA database. Each circle indicates a year where a recording is available for a speaker. Speaker names are shown on the y-axis, with their age in each recording given on the x-axis

Robust Speaker Verification for Biometric applications

Sponsor: IRCSET

Project description: Biometric verification involves the use of physiological features or behavioural traits of an individual to verify their identity. Examples of common biometric modalities are: fingerprints, the face, the iris, handwriting, gait and voice. Biometric systems have applications in securing access locally, e.g. airports, computer systems, restricted areas in a building, and remotely, e.g. over the phone or internet, along with additional applications in forensic investigation and information retrieval.

The advantage of voice as a biometric modality is that it can be obtained by a system in a way that is transparent to the user - this is in contrast to fingerprint and iris scans for example, which are often seen as invasive. In addition, voice can be transmitted without special equipment - it can be used to provide authentication over a normal telephone. Despite these advantages, using the voice in a biometric system presents many challenges. In a speaker verification system, a voice sample and a claimed identity are provided as input. The sample is compared to the user template in a database corresponding to the claimed identity, and based on a similarity score, the speaker is either verified or rejected. Achieving high accuracy speaker verification in real-world operating conditions is limited by variability between enrolment (creation of a speaker template) and verification (subsequent access of the system) sessions. The voice is subject to numerous natural sources of variability, from physiological sources like illness or ageing, to cognitive sources such as emotional state or speaking style. When the voice undergoes recording and transmission, further variability, due to background noise or the transmission channel, is introduced. The current focus of this project is on the effect of ageing on speaker verification. A new dataset, the Trinity College Dublin Speaker Ageing (TCDSA) database, has been compiled to investigate this effect. Our work in this area has demonstrated the significant deterioration of speaker verification accuracy due to ageing, and has proposed a new verification approach to address long-term variability due to both ageing and recording quality.

Page last modified on August 22, 2013