Forensic Phonetics: Speaker Identification in Legal Cases

This video introduces Automatic Speaker Recognition (ASR) systems, which are technologies designed to identify or verify speakers based on their voice characteristics. The system analyzes speech signals and extracts unique features from a speaker's voice to perform identification or verification tasks. ASR has become an essential aspect of voice biometrics with applications in authentication technologies.

ASR involves classification of: (1) Words - identifying individual words from speech, (2) Sentences - understanding sequences of words, and (3) Speaker identity - determining who is speaking. Classification is performed using statistical models and neural networks.

Speech scientists work on speech recognition (audio to text conversion), speech synthesis (text to speech with expressivity), and voice biometrics (speaker identification). Different applications require different training approaches for voice assistants, medical settings, or smartphone dictation. Speech synthesis has achieved natural-sounding voices but faces challenges in adding emotion and style. Voice biometrics and sentiment analysis can detect frustration for call center applications.

Machine learning is essential for developing voice biometric solutions. It involves training models using large datasets containing thousands of speakers, millions of utterances, dozens to hundreds of languages, and audio from various channels (landlines, mobiles, sat phones) and environments (noisy, quiet). This enables the technology to imitate human behavior through collected data.

Automatic Speech Recognition (ASR) is a field that transforms raw audio into corresponding word sequences through a pipeline involving audio capture, noise reduction, volume normalization, phoneme extraction (44 phonemes in English), and statistical probability analysis to deduce words and sentences. ASR systems use two main training approaches: human tuning (programmers adding commonly used words from conversation logs) and active learning (autonomous vocabulary expansion). The classical ASR pipeline integrates grammar, syntax, structure, and composition of audio signals, with applications including directed dialogue conversations (menu-based choices) and natural language conversations (open-ended interactions).

This section explores voice disguise as a voluntary change of voice features to conceal identity, with gender differences in disguise choices (men choose higher pitch, women choose lower pitch). The most frequently used disguises include falsetto, perturbed voice, and whisper. The section covers intoxication effects on speech, using the Exxon Valdez oil spill case where the captain's voice samples were analyzed to determine if he was under alcohol influence. Content identification focuses on determining what was said in audio recordings, particularly challenging with poor quality recordings. Filtering techniques remove background noise like air conditioning or fan sounds. The section also covers speaker profiling, which analyzes voice characteristics to determine regional background, physical characteristics, psychological traits, educational background, nationality, social status, and leadership qualities. Research shows dominant voices are associated with good leadership qualities.

The human vocal tract functions as a resonance chamber where air from lungs passes through vibrating vocal cords, producing excitation signals that resonate differently based on vocal tract shape. Different sounds emerge from changing tongue, jaw, and lip positions. Spectrograms reveal resonance patterns called formants. Key micro features include onset time (inertia-related delays between sounds), frequency modulation patterns, and harmonic structures. As people age, harmonics smear together and fundamental pitch lowers. Advanced AI techniques enable discovery of micro articulatory features using encoder-decoder architectures with reconstruction loss and generative adversarial networks. Voice disguise attempts fail because breath sounds cannot be fully controlled—the volume of air taken between sentences correlates with expected upcoming content, creating distinctive patterns. Pitch pulses also reveal individuality: separated pulses indicate normal speech while overlapping pulses suggest disguise. Expert mimics can change certain voice aspects but cannot alter others. Speaking rate changes push energy toward lower frequencies, disrupting consistent imitation since discriminatory information between sounds resides primarily in low frequencies.

The GSM voice channel operates at approximately 13 kilobits per second with a compression ratio of about 8:1. Input audio is typically 8 kHz sample rate with 13 significant bits, compressed down to 13 kbps. This high compression causes distortion for non-speech signals, requiring any data passed through the channel to resemble speech characteristics to avoid degradation.

This section covers speaker profiling and spectrograph analysis. Speaker profiling analyzes voice recordings to determine physical properties of the speaker, including sex, age, ethnicity, and geographic origin. Sex determination is most reliable because men typically have vocal frequencies between 60-180 Hz and women between 160-300 Hz, determined by vocal tract length. A 2012 study showed speaker profiling could identify a speaker's region with 85% accuracy. Forensic phoneticians compare voice recordings using spectrographs, which visualize acoustic properties. By examining where vowels are produced in the mouth and what consonants are used, analysts can determine whether two recordings are from the same speaker. This method focuses on articulatory features that are difficult to manipulate when someone is trying to disguise their voice.

SIM box calls experience audio degradation from three sources: source network digital transmission errors, VoIP packet loss during internet transmission, and over-the-air degradation from SIM box to tower. Cellular networks typically experience 4-5% loss rates due to the wireless medium. AMET focuses on detecting losses from the VoIP step, as these are distinct from legitimate over-the-air losses. For unconcealed losses, AMET uses short-term energy analysis to detect brief energy drops corresponding to lost packets. For concealed losses, AMET analyzes GSM packet loss concealment algorithms, which replay the last good packet at lower amplitude, creating echo patterns detectable through cepstral analysis. A peak at exactly 20 milliseconds in cepstral analysis indicates repeating audio, strongly suggesting GSM packet loss concealment.

Human ear witness testimony has fundamental limitations in forensic contexts. Witnesses often disagree with each other, and different levels of familiarity require separate evaluation frameworks. Unfamiliar voices are processed differently by the brain and should not be compared to familiar voices. Research shows recognition of unfamiliar voices decays rapidly—from 83% accuracy after one day to chance levels after five months. Expectations and cognitive biases significantly distort speaker recognition, as demonstrated by the Lindbergh kidnapping case and laboratory experiments showing skilled linguists failing to recognize clearly different voices due to expectation-driven perception.

The psychology of testimony reveals that witnesses, even those sincerely believing they are telling the truth, may suffer from cognitive biases and memory problems. These include forgetting, loss of memory, and false interpretations of events. The perception and memory through which knowledge of facts is acquired are subject to numerous problems that the psychology of testimony has identified. This undermines the persuasive approach to evidence valuation.

Ear witnesses are highly unreliable because different people hear different things from the same indistinct utterance. It is not sensible to ask experts 'what did you hear?' because answers will vary. Instead, experts should identify different hypotheses that people might hear. People can be easily primed to hear what they want to hear from indistinct utterances. In the Palestinian protest case, the video had 'gas the Jews' written on it, which primed viewers to hear that phrase. The more indistinct the utterance, the more powerful the priming effect. Experts should never tell people what they think was said.

Human memory is inherently unreliable and susceptible to cognitive biases, causing eyewitnesses to make false identifications approximately 30% of the time in controlled settings and at least 25% in real criminal cases; additionally, forensic evidence can be compromised when analysts are exposed to irrelevant contextual information, as demonstrated by studies showing fingerprint examiners changed their conclusions from '100% match' to 'no match' depending on the case context, highlighting why both eyewitness procedures and forensic analysis require strict scientific protocols to prevent contamination.

Auditory perception is a factor that can affect testimony credibility. In cases involving voice recognition, such as identifying criminal associates through voice, this factor becomes crucial. Voice can be considered like a fingerprint—unique and unchangeable through analysis. Legal operators can use this factor to establish credibility (a witness who recognized someone by voice despite other limitations) or to challenge it (a witness who cannot identify voices they have never heard before).