Mandi forensic lab to have auto voice, face comparison systems

Tribune News Service

Mandi, September 4

The Regional Forensic Science Laboratory (RFSL) in Mandi is developing automatic voice and face comparison systems that are likely to be put to use soon. The systems have been developed from the Innovation Fund of the state government. The project is being executed by Dr Rajesh Verma, an official in the RFSL, in collaboration with IIT-Mandi.

Rajesh Verma said, “Voice samples are traditionally taken from the suspect and contains the same words said in the same way to be compared with the suspect recording. This sometimes creates problem for comparison if the speaker changes his speaking style.”

“The comparison process is also time-consuming and requires a lot of efforts in segregating speakers and words for comparison. Automatic voice comparison is a text independent approach that does not require the speaker to utter the same words in the same way,” he added.

He said, “Face comparison from CCTV footage and the suspect is also a problematic area, where low-quality images from the CCTV footage make the facial comparison difficult, especially when the pose of the two faces is different. The face shape and other gross features of the face can be compared in such situations.”

“The automatic system being tested at the RFSL is based on deep learning techniques where the algorithm has been trained to detect the facial features on millions of images. The system predicts the facial features of the questioned face and automatically compares it with the suspect’s face even if the pose of the two faces are different. The test results in trials have reported over 95 per cent accuracy in identification of the correct person from images with varying poses,” he said.

Testing and validation of both methods are going on and reference databases will be created for voices and faces from representative populations. The current internationally accepted trend in reporting such comparisons is the likelihood ratio approach and Dr Rajesh Verma has done some pioneering work on adoption of this approach.

Time-consuming process

• Voice samples are traditionally taken from the suspect and contains the same words said in the same way to be compared with the suspect recording
• This sometimes creates problem for comparison if the speaker changes his speaking style
• The comparison process is also time-consuming and requires a lot of efforts in segregating speakers and words for comparison
• Automatic voice comparison is a text independent approach that does not require the speaker to utter the same words in the same way