语音科学与数字信号处理基础

Human speech, a seemingly effortless ability for most, represents one of the most complex and nuanced signal processing tasks encountered in nature. It is a continuous stream of acoustic waves, shaped by a intricate biological apparatus, carrying meaning through a structured linguistic code. The journey from the speaker's intent to the listener's understanding involves a cascade of physical, physiological, and cognitive processes. For machines to participate in this communication—to recognize what is said or to synthesize speech that sounds natural—we must first demystify this cascade. This chapter serves as that essential foundation, constructing a multidisciplinary bridge spanning anatomy, physics, signal processing, and linguistics. We begin by examining the biological instrument of speech: the human vocal tract. We will explore how the lungs provide the power source, the vocal folds generate a periodic or aperiodic sound source, and the supralaryngeal vocal tract—the throat, mouth, and nasal cavities—acts as a dynamic filter, imprinting the unique spectral signatures that distinguish one sound from another. This source-filter model is not merely an analogy; it is the cornerstone of modern speech analysis, providing a parametric framework that separates the fundamental frequency (related to pitch) from the resonant frequencies (formants) that define vowel quality and consonant place of articulation.

Following this physiological grounding, we transition to the mathematical representation of speech. Raw acoustic pressure waves are analog, continuous phenomena. To be analyzed by digital computers, they must undergo discretization—both in time (sampling) and amplitude (quantization). The Nyquist-Shannon sampling theorem establishes the critical relationship between the sampling rate and the highest frequency present in the signal, a principle that dictates the design of all digital audio systems. Once digitized, the challenge becomes extracting meaningful, compact features from this high-dimensional time-series data. This is the domain of time-frequency analysis. We will delve into the Fourier Transform (FT), the mathematical engine that decomposes a signal into its constituent sinusoidal frequencies. However, speech is non-stationary; its spectral characteristics change rapidly over time. The Short-Time Fourier Transform (STFT) addresses this by applying the FT to short, overlapping windows of speech, yielding a spectrogram—a visual representation of how energy distributes across frequencies as time progresses. To better align with human auditory perception and to reduce feature dimensionality, we introduce Mel-frequency analysis. The Mel scale is a perceptual pitch scale where equal distances correspond to equal perceived pitch differences. Computing filterbank energies on a Mel scale (MFCCs—Mel-Frequency Cepstral Coefficients) produces a feature vector that is both compact and robust, forming the de facto standard for decades of speech recognition systems.

Finally, we must connect these physical and acoustic representations to the abstract symbolic system they convey: language. A brief introduction to phonetics and phonology is indispensable. We define the phoneme as the smallest unit of sound that can distinguish meaning (e.g., /b/ vs. /p/ in "bat" vs. "pat"), and the allophone as context-dependent variants of a phoneme. Concepts like voicing, place of articulation, and manner of articulation provide the descriptive vocabulary for consonant sounds, while vowel height and backness define vowel spaces. The syllable structure offers a intermediate rhythmic and organizational unit. Understanding these linguistic categories is crucial because the goal of automatic speech recognition is to map a continuous acoustic signal onto a discrete sequence of these linguistic units, and the goal of synthesis is to generate acoustic waveforms that correctly instantiate them.

This chapter, therefore, equips the reader with a tripartite lens: the biological source of speech, the mathematical tools for its digital representation, and the linguistic targets it encodes. Mastery of these fundamentals is not an academic exercise but a practical necessity. The feature extraction techniques described here directly feed the acoustic models of Chapters 3 and 4. The source-filter paradigm informs the generative models of hidden Markov models and, later, the architectural inductive biases of neural vocoders. The phonological distinctions define the output label space for recognition systems. As we progress from the deterministic, physics-based models of this chapter to the statistical and then end-to-end neural models of later chapters, remember that all these algorithms are ultimately trying to model, approximate, or bypass the complex, layered process outlined here. The subsequent evolution in speech technology can be seen as a continuous effort to find more efficient and accurate mathematical surrogates for this intricate human chain of production and perception.

1.1 The Human Speech Production Mechanism: A Source-Filter Perspective

The generation of speech sounds is a remarkable biomechanical process. Air from the lungs is converted into a rich spectrum of sounds by the coordinated action of the larynx and the vocal tract. The most successful and enduring model for describing this process is the source-filter theory of speech production, independently proposed by Gunnar Fant in 1960 and building on earlier work by Chiba and Kajiyama. This model elegantly separates the speech signal into two largely independent components: the excitation source and the acoustic filter.

1.1.1 The Excitation Source: The Larynx and Vocal Folds

The power source for most speech sounds originates in the lungs. During exhalation for speech, the diaphragm and intercostal muscles maintain a stable subglottal pressure (typically between 5 and 20 cm H₂O above atmospheric pressure). This pressurized air flows up the trachea to the larynx, often called the voice box, situated at the top of the trachea.

Within the larynx reside the vocal folds (or vocal cords), two bands of muscular tissue stretched horizontally across the laryngeal cavity. Their position and tension are controlled by intrinsic laryngeal muscles. The space between them is the glottis.

The manner in which the vocal folds vibrate determines the nature of the excitation source:

• **Voiced Sounds (Periodic Excitation):** When the vocal folds are brought close together and set into tension, the increasing subglottal pressure forces them apart (abduction). As they separate, the pressure drops, and the Bernoulli effect, combined with tissue elasticity, pulls them back together (adduction). This cycle repeats, generating a periodic train of air pulses. The rate of this vibration is the **fundamental frequency (F0)**, perceived as pitch. F0 is influenced by the length, tension, and mass of the vocal folds, which are in turn controlled by the cricothyroid and thyroarytenoid muscles. Typical F0 ranges are 80-150 Hz for adult males, 150-250 Hz for adult females, and 250-400 Hz for children.
• **Voiceless Sounds (Aperiodic/Noise-like Excitation):** If the vocal folds are held apart sufficiently, the air flows through the glottis without causing self-sustained oscillation. The resulting excitation is turbulent, noise-like airflow, generated by the constriction and high-velocity jet of air. This is the source for sounds like /s/, /f/, /θ/ (as in "thin").
• **Glottal Stops:** A complete closure of the vocal folds, building pressure that is released explosively, as in the catch in "uh-oh."

The waveform produced at the glottis, called the glottal flow, is not a simple periodic square wave. Due to the physics of the flow and the tissue dynamics, it has a characteristic shape with a rapid opening phase and a slower closin