Reading Spectrograms: Consonant Classification Explained

Sociophonetics is the interdisciplinary study of how speech production and perception reveal and construct personal speaker-related identities, including emotional states, social affiliations, and group memberships; this field examines how linguistic variation across dialects, gendered varieties, and ethnic varieties provides insights into language structure and social meaning, moving beyond traditional focus on standard languages to understand how speech reflects complex layers of individual and collective identity.

Praat is the most widely used software for acoustic analysis in phonetics, developed in 1991 by David van den Berg. It provides tools for recording, analyzing, and visualizing speech data. Key features include automatic speech activity detection, formant analysis, and fundamental frequency extraction. Automatic speech detection may not accurately capture word onsets for obstruents (p, b, t, d), requiring manual adjustment. For clinical populations, automatic analysis may be less reliable. The software allows extraction of duration, pitch, and intensity measurements, with Praat scripts enabling semi-automatic data extraction from annotated recordings.

Praat is a free software program that converts audio signals into visual representations (waveforms and spectrograms), enabling linguists and clinicians to measure acoustic properties like pitch, loudness, and formants to analyze speech patterns, identify voice disorders, and study phonetic variations across languages.

Acoustic phonetics examines speech transmission through sound waves and their physical properties. Quality recordings require minimal background noise, specialized external microphones, and controlled laboratory conditions. Computer-aided analysis emerged in the 1960s, enabling sophisticated research. Major software tools include Praat (spectral, formant, and pitch analysis), WaveSurfer (open-source visualization), Speech Analyzer (multi-level transcription), and SFS (University College London's research platform). Voice Onset Time (VOT) measures stop burst release timing, classifying stops into aspiration categories. Chiao and Chen's 2008 study compared Mandarin and English aspirated stops, finding Mandarin stops occupied higher and wider VOT ranges, supporting four-category classification systems over three-category models.

Forensic phonetics is the application of phonetic principles to criminal investigations, focusing on speaker identification, audio authentication, and voice analysis. Each human voice is unique due to anatomical factors (vocal tract length, oral and nasal cavity size), physiological factors (vocal fold vibration frequency - approximately 120 Hz for men, 240 Hz for women, 330 Hz for children), and environmental factors (language exposure, family background). Key applications include signal analysis (examining acoustic signatures of gunshots), speaker identification (comparing criminal speech samples with suspect samples), and speaker profiling (determining gender, age, regional background, and psychological characteristics from voice features). Challenges include voice disguise, intoxication effects, and the emergence of voice cloning technology. The field evolved from cases like the Timothy Evans case (1949-1953) and Lindbergh kidnapping (1932), which highlighted the importance of scientific voice analysis in criminal justice.

This extensive section covers forensic phonetics, voice analysis, speaker profiling, and authorship attribution. Forensic phoneticians analyze speech and audio evidence, helping police conduct voice lineups where witnesses identify suspects based on voice recordings. Linguists analyze recordings for similar characteristics in pitch, intensity, and formant frequencies to find matching voices. The David Bain case demonstrates that prior knowledge and expectations (priming) can dramatically affect what people hear in audio recordings—when 200 participants listened to the same audio without knowing the case details, none heard the alleged confession. Pronunciation variations are explored through the 'can' vs. 'can't' case, where forensic speech scientists used spectrograph analysis to prove a doctor said 'can't' rather than 'can.' Speaker profiling analyzes voice recordings to determine sex, age, ethnicity, and geographic origin, with sex determination being most reliable (men: 60-180 Hz, women: 160-300 Hz). Authorship attribution identifies who wrote a text by analyzing linguistic features, demonstrated through the Robert Galbraith case where computational linguists revealed Galbraith was likely a pseudonym for JK Rowling. Stylistic analysis examines unique patterns in how individuals use language (idiolect), such as preferences for sentence structures with ditransitive verbs. In the JonBenet Ramsey case, forensic linguists analyzed a ransom note examining misspellings, word choices, and morphological patterns to clear the parents from suspicion. In the Jenny Nichol case, linguists compared specific linguistic patterns to prove that murderer David Hodgson had sent texts pretending to be Nichol.

Forensic phonetics has grown significantly, using laboratory instruments to identify speakers by comparing voice samples. Speech technology works on three fronts: synthesis (computers vocalizing text), recognition (machines understanding speech), and interaction (human-machine dialogue). These systems power applications like voice-to-text in cell phones and automated telephone services. Translation professionals need knowledge of sound systems in both source and target languages to ensure accurate communication. Understanding phonetics and phonology is essential across multiple professional fields.

This comprehensive introduction covers the foundational concepts of forensic phonetics, the application of linguistic knowledge to legal contexts for voice evidence analysis. The lecture explains that voice evidence includes recorded sounds from crime scenes and suspect recordings, analyzed using sound spectrograms to create 'voice prints.' Three core applications are discussed: detecting voice manipulation, speaker profiling to identify demographic characteristics, and speaker comparison to determine if two voice samples come from the same person. The analysis reveals multiple information layers: phonemic information (speech sounds and meanings), physiological information (body size, gender, age), sociolinguistic information (social class, education, dialect), and individual-specific characteristics. The most challenging application is speaker comparison, which requires rigorous scientific methodology beyond simple listening.

Alpha ratio analysis shows negative values where more negative indicates steeper spectral slope and higher dysphonia probability. Voice Report provides jitter (normal <1%) and shimmer (normal <3%) measurements. Spectrogram analysis reveals fundamental frequency (women: 190-262 Hz), intensity (60-80 dB), and formant patterns. F1 correlates with vocal opening (higher = more open), F2 with vowel anteriority (higher = more anterior). Harmonic analysis shows resonance patterns through dark/light zones. Voice 6 showed normal F0 (240 Hz) while Voice 1 exceeded normal range (396 Hz). Acoustic analysis is essential for voice disorder diagnosis and rehabilitation in speech-language pathology.

This segment covers the clinical framework for speech sound disorders: (1) Evaluation and screening using specific filters and tools to identify the true origin of phonological alterations; (2) Determining when referral or intervention on other aspects is necessary; (3) Contextual elements and specific questions for clinical assessment; (4) Clinical diagnosis using updated clinical nomenclature; (5) Qualitative and objective diagnosis documentation; (6) Clarifying the speech-language pathologist's role according to diagnosis and treatment approach; (7) Strategic treatment planning based on phonetic repertoire analysis according to patient age. The presenter also covers treatment planning: organizing an intervention route with a five-step process for adapting praxias objectively and functionally, installing correct articulatory patterns, and providing ready-to-use tools for the three stages of intervention (conscientization, production, and transfer).

A comprehensive voice assessment involves gathering patient history, conducting perceptual evaluations using tools like the GRBAS scale, performing oral motor examinations, measuring maximum phonation time and S/Z ratio, and conducting acoustic analysis using Praat software to record and analyze pitch glides, habitual pitch, intensity, and perturbation measures (jitter, shimmer, noise-to-harmonics ratio) to establish baseline objective measures for tracking voice disorder progression and treatment outcomes.

Proper diagnosis requires assessment of phonetic and phonological production, detailed history-taking, and orofacial morphological examination. The non-symptomological diagnosis (DSM-5 based) is straightforward, while symptomological diagnosis requires detailed error pattern analysis. Intervention focuses on phonetic and phonological alterations: phonetic errors use proprioceptive cues and phonetic guidance, while phonological errors require addressing cognitive-linguistic levels beyond simple discrimination tasks.

Modern voice assessment requires specific equipment: microphone with flat frequency response, external sound card, and proper recording setup. The VoxMetria 5.0 system provides user-friendly extraction of CPP, CPS, and other measures from sustained vowels and connected speech. CPP measures how well harmonic energy stands out from noise, while CPS measures the cepstral peak slope. The system allows comparison of detailed measures (CPS) versus global measures (CPP). Acoustic analysis is essential because it makes the abstract phenomenon of voice concrete and measurable, providing the best documentation under standardized conditions.

The acoustic model in automatic speech recognition uses Hidden Markov Models (HMM) to convert extracted speech features (such as MFCCs, delta MFCCs, and delta delta MFCCs) into sequences of hypothesized phones (the basic units of speech). The HMM consists of an emission model (probability of observing features given a phone state) and a transition model (probability of transitioning between phone states). The Viterbi algorithm is then applied to find the most likely sequence of hidden states (phones) that could have produced the observed feature vectors, enabling the system to recognize spoken language.

Acoustic models map feature vectors to phoneme sequences using probabilistic frameworks. Hidden Markov Models use states representing phonemes with transition and emission probabilities learned from training data. Deep Neural Networks provide posterior probabilities over phonemes for each frame. Pronunciation models link phoneme sequences to words using expert-derived dictionaries, though real speech exhibits extensive variation beyond dictionary pronunciations, with words averaging 4-5 pronunciations in spontaneous speech due to accents, speaking speed, and individual patterns. Articulatory phonology offers an alternative representation using multiple streams of vocal tract variables, explaining variation through asynchrony between articulatory features.

Acoustic modeling in speech recognition involves learning emission probabilities that map between hidden states (sounds) and observed acoustic features. Each sound has a characteristic distribution over frequency-domain features extracted from speech signals (typically 39-dimensional vectors). These features are extracted by taking 25ms windows of speech, applying Fourier transforms, and extracting coefficients. The model learns what acoustic patterns correspond to each phonetic unit, enabling the system to recognize which sounds are being produced from the observed acoustic signal.

Acoustic modeling is the process of converting raw speech waveforms into sequences of phonetic units (phones). The standard approach involves: (1) dividing the long speech waveform into small overlapping frames (typically 10ms), (2) extracting features from each frame, and (3) using a classifier to determine which context-dependent phone state corresponds to that frame. This transforms continuous audio signals into discrete phonetic representations that can be used for speech recognition.

Acoustic phonetics has important applications in technology and healthcare: (1) Speech recognition - acoustic features such as formants, pitch, and waveforms are critical for identifying spoken words. (2) Speaker identification - unique characteristics of a speaker's voice, including speech formants and harmonics, enable identification of individuals. (3) Text-to-speech synthesis - acoustic phonetics helps synthesize natural-sounding speech using spectrograms and formants. (4) Speech therapy - acoustic analysis is used to diagnose and treat speech disorders.