News

Oxford Wave Research are delighted to be collaborators with the University of York and the Netherlands Forensic Institute in a recently awarded ESRC-funded project (£1,012,570) ‘Person-specific automatic speaker recognition: understanding the behaviour of individuals for applications of ASR’ (ES/W001241/1). This is a three year project running from 2022 to 2025 led by Dr Vincent Hughes (PI), Professor Paul Foulkes (CI) and Dr Philip Harrison in the Department of Language and Linguistic Science at the University of York. The project is due to start in summer 2022 and will run for 3 years. OWR will be providing our expertise and consultancy in automatic speaker recognition and our flagship VOCALISE forensic speaker recognition system. 

Automatic speaker recognition (ASR) software processes and analyses speech to inform decisions about whether two voices belong to the same or different individuals. Such technology is becoming an increasingly important part of our lives; used as a security measure when gaining access to personal accounts (e.g. banks), or as a means of tailoring content to a specific person on smart devices. Around the world, such systems are commonly used for investigative and forensic purposes, to analyse recordings of criminal voices where identity is unknown. 

The aim of this project is to systematically analyse the factors that make individuals easy or difficult to recognise within automatic speaker recognition systems. By understanding these factors, we can better predict which speakers are likely to be problematic, tailor systems to those individuals, and ultimately improve overall accuracy and performance. The project will use innovative methods and large-scale data, uniting expertise from linguistics, speech technology, and forensic speech analysis, from the academic, professional, and commercial sectors. This has been made possible via the University of York’s strong collaboration with Oxford Wave Research and two project partners including the Netherlands Forensic Institute (NFI).

The University of York and OWR teams are looking forward to a very fruitful collaboration that will undoubtedly further the state of the art in forensic speaker recognition.

Dr Vincent Hughes, Principal Investigator, University of York says “We are delighted to be working so closely on this project with Oxford Wave Research, who are world leaders in the field of automatic speaker recognition and speech technology. We hope that our research will deliver major benefits to the fields of speaker recognition and forensic speech science”.

Dr Anil Alexander, CEO, Oxford Wave Research says “Our team led by Dr Finnian Kelly is thrilled to contribute to this in-depth study of the individual specific-factors affecting speaker recognition, with the accomplished research team led by Dr Hughes from the University of York who are at the forefront of this space, and real-word end-users like the Netherlands Forensic Institute who have been driving research and innovation in this space for many years”.

Taking this idea a step further, alternate reality games (ARGs) are a modern spin on the traditional scavenger hunt in which participants scour websites, social media, and videos looking for clues. These games have taken social media sites like YouTube and Reddit by storm over the last few years. Organisers, often video game developers, will bury information in all sorts of places, like images, website code, and audio. Whole communities have been formed in order to find out secret stories and previews for their favourite video game, or just to have some cooperative fun while solving a digital mystery.

Epic Games created ARG content in the run-up to the Season 5 release of their famous multiplayer online game, Fortnite Battle Royale. They staged a rocket launch within the video game itself, during which some of the audio played was slightly odd. Gamers quickly realised that there was probably more to the audio clip than what could be just heard. Looking for patterns within the audio led them to visualising the frequencies in the audio in a spectrogram.  One such example of using SpectrumView to analyse the audio clip by player “Rockin Thomas86” is shown in the video below.

On the spectrogram, you can see pixelated skulls at the start and end of the audio, and, in the middle, a list of letters and numbers. According to the Game Detectives Wiki, the skull shapes were shown on television screens within the game before the rocket launch, while the letters and numbers could be decoded as ASCII values to produce in-game coordinates. Some time after the rocket launch, dimensional rifts opened up at these coordinates, causing locations to appear and disappear on the game map. Players were primed to check the locations, having teased out the message hidden in the rocket launch audio.

Spectrum analysers like our iOS app SpectrumView can open up a whole new dimension of information in audio, and we are excited to see what more our users can find hidden away in the audio of all sorts of ARG content.

Virtual Event, Virtual Stand, even Virtual Sweets, but the same real people!

The OWR team, Nikki, Ekrem, Oscar and Anil warmly welcome you to join us at the first ever virtual Security & Policing event taking place 9-11 March 2020.

This year’s event coincides with our 10 year anniversary and we are proud to be sharing with you our state-of-the-art desktop and ‘on-device’ speaker recognition and audio processing software in use at the forefront of the voice biometric field.

Two of our collaborative papers, one on voice spoofing detection, and the other on the effects of device variability on forensic speaker comparison, are appearing at this week’s virtual ODYSSEY 2020 Speaker and Language Recognition Workshop. Video presentations for both papers are now available on the workshop website: http://www.odyssey2020.org/

The full papers, along with the rest of the conference proceedings, can be found at: https://www.isca-speech.org/archive/odyssey_2020/index.html

In our paper with Bence Halpern (PhD student, University of Amsterdam), “Residual networks for resisting noise: analysis of an embeddings-based spoofing countermeasure,” we propose a new embeddings-based method of spoofed speech detection using Constant Q-Transform (CQT) features and a Dilated ResNet Deep Neural Network (DNN) architecture. The novel CQT-GMM-DNN approach, which uses the DNN embeddings with a Gaussian Mixed Model (GMM) classifier, performs favourably compared to the baseline system in both clean and noisy conditions. We also present some ‘explainable audio’ results, which provide insight into the information the DNN exploits for decision-making. This study shows that reliable detection of spoofed speech is increasingly possible, even in the presence of noise.

See a blog post from Bence (including some explainable audio examples) here: https://karkirowle.github.io/publication/odyssey-2020

In our paper with David van der Vloed (from the Netherlands Forensic Institute), “Exploring the effects of device variability on forensic speaker comparison using VOCALISE and NFI-FRIDA, a forensically realistic database,” we investigate the effect of recording device mismatch on forensic speaker comparison with VOCALISE. Using the forensically-realistic NFI-FRIDA database, consisting of speech simultaneously-recorded on multiple devices (e.g. close-mic, far-mic, and telephone intercept, as seen in the data collection image), we demonstrate that while optimal performance is achieved by matching the relevant population recording device to the case data recording device, it is not necessary to match the precise device; broadly matching the device type is sufficient. This study presents a research methodology for how a forensic practitioner can corroborate their subjective judgment of the ‘representativeness’ of the relevant population in forensic speaker comparison casework.

In these recent months of 2020, like many others around the world, we have found ourselves adjusting to the new normal of wearing masks in various places like supermarkets and other public spaces. We found ourselves (minorly) annoyed that some biometric identification, like face recognition, doesn’t quite work when wearing masks. This made us wonder how well voice biometric solutions could work when speakers are wearing masks, and we decided to perform a small experiment to analyse this. 

Over the last few weeks, we have been performing some small-scale tests of our VOCALISE and PHONATE software against speech spoken from behind a mask. We have found our systems to be quite robust to masked speech – they are able to recognise speakers across different mask-wearing conditions well.

The video below explains our experiment and discusses our findings. We hope that you find it interesting! 

Similar-sounding voices is of interest in many areas, be it for voice parades in a forensic setting, voice casting for film-dubbing or voice banking to save one’s voice for future synthesis in case of a degenerative disease. However, it is a very time-consuming and expensive task. With the aim of finding an objective method that could speed up the process, we considered an automatic approach to rate voice similarity and explored the relationship between voice similarity ratings made by a total of 106 human listeners – some of whom may have been you – and comparison scores produced by an i-vector-based automatic speaker recognition system that extracts perceptually-relevant phonetic features. Our results showed a significant positive correlation between human and machine, motivating us to continue our developments in this space.

The main highlights of this work are that human judgements of voice similarity are seen to correlate with automatic speaker recognition  assessments (using auto-phonetic features) (this trend was seen with both English and German speakers’ judgements of English voices). These automatic speaker recognition assessments therefore show potential for automatically selecting foil voices for voice parades.

This paper is based on Linda’s Gerlach’s master’s thesis work (University of Marburg, Germany) at Oxford Wave Research last year and uses the phonetic mode of VOCALISE speaker recognition software.

The full paper is available for free download on the Journal’s webpage. Please check the following link for the full abstract and paper, available for free using this link before 19th November 2020:

https://authors.elsevier.com/a/1bqZu_3pyeDhKh