December 3, 2021
By Pavel Jirik in Blog
Forensic experts play an important role during crime investigations. Their unique knowledge of this specific subject combined with the meticulous examination of key details can often add further clarity to the case being investigated and provide crucial evidence for courts and law enforcement agencies.
Forensic experts not only rely on their advanced knowledge and attention to detail but also on cutting-edge forensic software, which helps them provide accurate results as fast as possible.
One of such challenging forensic disciplines is voice comparison.
Determining whether the speaker’s voice is the same in two or more audio recordings requires careful consideration of many voice properties, which can be a very time-demanding task.
And with the rise of automatic speaker recognition systems powered by deep neural networks, it was only a matter of time before a standardized forensic evaluation method would be defined to evaluate the accuracy of forensic voice comparison software.
Forensic_Eval_01 – The Evaluation of Forensic Voice Comparison Systems
Aware of the increasing pressure on forensic laboratories to validate the performance of forensic voice comparison systems in order to assess the strength of evidence presented in court, Geoffrey Morrison and Ewald Enzinger introduced an evaluation method called forensic_eval_01 in 2016.
The full name of their evaluation approach is the multi-laboratory evaluation of forensic voice comparison systems under conditions reflecting those of a real forensic case.
As the name suggests, the evaluation is designed to test voice comparison systems under the conditions of real-life forensic cases, considering various forensic challenges such as poor recording quality and a different speaking style between the known- and questionable-speaker recordings.
Furthermore, it enables a clear comparison of the speaker identification systems available on the market with each other.
The 2019 Forensic Evaluation of Phonexia Speaker Identification
In 2019, Phonexia Speaker Identification technology was tested by forensic experts from Bundeskriminalamt (the Federal Criminal Police Office of Germany) in collaboration with the Israeli police based on forensic_eval_01 methodology.
As we described in our blog post at the time, Phonexia Speaker Identification delivered an outstanding performance, achieving an Equal Error Rate (EER) of 2.2% after calibration.
In comparison to the other speaker recognition systems that were being tested based on forensic_eval_01 methodology as well, Phonexia Speaker Identification provided the most accurate voice comparison performance out of all of them.
It outperformed voice comparison systems such as Nuance Forensics 11.1 (3.1% EER) and BatVox 4.1 (9.6% EER).
The 2021 Forensic Evaluation of Phonexia Speaker Identification
Fast forward to the present day, and our latest commercially available generation of Phonexia Speaker Identification was tested based on the forensic_eval_01 method once again.
This time by Andrea Fröhlich, a forensic expert from the Zurich Forensic Science Institute in Switzerland.
She used our latest generation of deep neural networks for speaker recognition, which offers not only exceptional accuracy out of the box but also comes with a built-in, easy-to-use calibration feature, allowing seamless, state-of-the-art adaptation (mean normalization) of the system’s accuracy based on a given channel.
She achieved the following forensic_eval_01 results:
- 2.1% EER before calibration
- 1.2% EER after calibration
As you can see, the latest generation of Phonexia Speaker Identification reduced its Equal Error Rate by almost half compared to the results from 2019!
Phonexia Offers the Most Accurate Voice Comparison Technology Currently Available on the Market
The extremely low 1.2% EER is a fantastic result and massive improvement of Phonexia Speaker Identification's performance over the previous forensic_eval_01 evaluation.
It also means that, based on forensic_eval_01 evaluations performed on other voice comparison systems up until now, Phonexia currently provides the world’s most accurate, commercially available technology for forensic voice comparisons.
And because our research and development teams never stop pushing the capabilities of our speech technologies further, we are looking forward to how our solutions will perform even better in the upcoming forensic evaluations.
Would you like to leverage our speaker identification technology in your projects?