What Is Speech?

Speech, as a subject of scientific and engineering study, is the acoustic signal produced by the human vocal apparatus for the purpose of linguistic communication. It is studied across disciplines including acoustics, linguistics, physiology, signal processing, and artificial intelligence, each of which focuses on a different level of abstraction: from the biomechanical movements of the articulators to the abstract phonemes, words, and syntactic structures that the acoustic signal encodes. Engineering applications of speech science range from telephone bandwidth optimization to voice-controlled interfaces and clinical diagnosis of vocal pathologies.

The field draws its analytical foundations from two primary disciplines: acoustic phonetics, which characterizes the physical properties of speech sounds, and digital signal processing, which provides the mathematical tools for analyzing, compressing, and recognizing those sounds. IEEE has documented the evolution of speech technology through the IEEE Transactions on Acoustics, Speech, and Signal Processing, published from 1974, and its successors, which collectively constitute the primary research record for the field.

Speech Production and Acoustics

Speech is produced by a cascade of physiological processes. Air expelled from the lungs passes through the larynx, where the vocal folds vibrate to generate a periodic excitation signal whose fundamental frequency, denoted F0, is perceived as voice pitch. This glottal source signal then propagates through the vocal tract, which acts as a resonant acoustic filter. The resonant peaks of the vocal tract transfer function are called formants. The first two formant frequencies, F1 and F2, are primarily responsible for distinguishing vowel sounds: F1 correlates inversely with vowel height and F2 with tongue frontness, creating a two-dimensional vowel space used in acoustic phonetics to classify all vowel sounds across human languages. Consonants involve additional articulatory gestures including complete closures of the vocal tract (stops), partial constrictions (fricatives), and nasal coupling. The source-filter theory of speech production, as detailed in materials from MIT's Linguistic Phonetics course, provides the mathematical framework connecting articulatory configurations to the acoustic output.

Speech Processing and Representation

In digital systems, speech signals are captured, sampled, and represented as sequences of acoustic feature vectors for analysis, coding, or recognition. Linear predictive coding (LPC) models the vocal tract as an all-pole filter and estimates its coefficients from short frames of speech, producing a compact spectral representation widely used in low-bitrate voice codecs. Mel-frequency cepstral coefficients (MFCCs), computed from the log mel-scale spectrum of short speech frames, capture the spectral shape of speech in a form that correlates well with human auditory sensitivity and serves as the standard feature representation for automatic speech recognition (ASR) systems. Deep neural networks applied to these features, or directly to raw waveforms, now achieve word error rates below five percent on standard benchmark tasks in quiet conditions. The IEEE Signal Processing Society's IEEE/ACM Transactions on Audio, Speech, and Language Processing publishes current research spanning speech enhancement, coding, recognition, and synthesis.

Speaker Recognition

Speaker recognition refers to the automatic identification or verification of a person's identity from their voice. It exploits speaker-specific characteristics of speech that arise from anatomical differences in vocal tract length, vocal fold mass, and habitual articulation patterns. Text-dependent speaker verification requires the speaker to say a specific pass-phrase, allowing the system to match both phonetic content and speaker characteristics. Text-independent recognition operates on arbitrary speech content and must model speaker identity across varying linguistic contexts. Modern speaker recognition systems use i-vectors or x-vectors as compact embeddings of speaker identity extracted by deep neural networks, achieving equal error rates well below one percent in controlled conditions. Forensic applications require additional caution because acoustic conditions, emotional state, and vocal health can alter the voice. A foundational survey of speaker recognition methods is available through the Springer chapter on speaker recognition, covering both classical and statistical approaches.

Applications

Speech has applications in a range of fields, including:

• Voice-controlled consumer devices and smart speakers using automatic speech recognition
• Telephone network voice coding and transmission using low-bitrate speech codecs
• Clinical diagnosis and monitoring of speech disorders and neurological conditions
• Forensic speaker identification in law enforcement and legal proceedings
• Accessibility technologies including real-time captioning and voice synthesis for augmentative communication