Designing an IVR-based Speech Data Collection Framework for building Realistic Speech Corpus on Forensic Automatic Speaker Recognition

Automatic speaker recognition in forensic scenarios or Forensic Speaker Recognition (FSR) works involve serious challenges with respect to highly case-specific audio data processing, handling mismatched voice conditions for variability in device, language, speaking style and background noise processing. The dearth of essential data resources, useful research studies and proper implementation strategy on FSR, so far, have hindered the research and development of reliable automated FSR methods that can be applied to solving real-life audio forensics cases. The automated approach will also minimize sole dependency on the age-old expert-guided auditory-acoustic-based voice comparison methods. This paper presents a brief informative review of the existing FSR approaches, datasets and toolkits and highlights some dominant speaking scenarios under the audio forensics domain. An Interactive Voice Response (IVR) based automated speech data collection framework has been indigenously developed based on this review. The IVR call flow steps, recording content, data and metadata specification, data naming and storage details and data collection guidelines have been presented with a clear, descriptive design diagram of the IVR call flow. Our aim is to build a forensic speech corpus considering the population, spoken language, demographic and socio-cultural contexts and support the community to experiment and evolve better methods for automated FSR.