Using Nuance Forensics

Nuance Forensics is a web-based application built on the Nuance Security Suite. Its general purpose is to statistically assess whether unknown audio segments match target speakers.

Nuance Forensics provides dedicated tools for managing entities that include forensic assessments, audio segments, speaker profiles, and reference populations. Investigators can calculate likelihood ratios (LRs, the standard reporting mechanism for forensic scientists), and can use Tippett plots for indications of LR reliability (see Creating a Tippett plot). The system produces printable forensic reports and zip archives for each assessment.

To use the tool, operators load audio into the system database, associate identities with the audio, and process files to create voiceprints. Investigators can search the database for audio, and make comparisons. At any time, new recordings can be added, processed, compared, and analyzed. For more details, see Workflows (processes and procedures).

This figure shows key concepts and components used when comparing audio:

The questioned sample is the audio recording under analysis. For example, the unknown voice whose identity you wish to validate.          The reference population is similar to a line-up of suspects with homogeneous appearances to avoid false positive matches.      It is a set of recordings with language characteristics that are similar to the suspect speaker sample      (same gender, language, accent, and dialect). The tool provides dozens of generic reference populations.       The suspect speaker is one or more of a speaker’s voiceprints selected from the database.           The background model normalizes signaling differences in the recordings      (for example, when the audio for questioned and suspect voiceprints are collected from different audio channels).         To summarize, the figure shows a sample being compared to a collection of voiceprints.      The system uses a background model to remove signaling differences between the samples.      The sample is also compared to a reference population to avoid false positive matches.

Explanation of the figure:

  • The questioned sample is the audio recording under analysis. For example, the unknown voice whose identity you wish to validate.
  • The reference population is similar to a line-up of suspects with homogeneous appearances to avoid false positive matches. It is a set of recordings with language characteristics that are similar to the suspect speaker sample (same gender, language, accent, and dialect). The tool provides dozens of generic reference populations. See Using the Reference Populations Management View.
  • The suspect speaker is one or more of a speaker’s voiceprints selected from the database.
  • The background model normalizes signaling differences in the recordings (for example, when the audio for questioned and suspect voiceprints are collected from different audio channels).
  • To summarize, the figure shows a sample being compared to a collection of voiceprints. The system uses a background model to remove signaling differences between the samples. The sample is also compared to a reference population to avoid false positive matches.