Why Voice Models Become Embodied

How high-fidelity speech synthesis converges on an embodied simulation

A powerful waveform model does not need a literal anatomical simulator with variables named lung_volume, tongue_position, or thyroarytenoid_activation. It can learn shortcuts, and for ordinary speech those shortcuts may work remarkably well.

They become less reliable as the demands increase. A model may be asked to remain coherent across long utterances; respond to unusual instructions or physical perturbations; preserve one speaker’s physiology from scene to scene; produce combinations absent from its training data; distinguish similar sounds with different causes; interact in real time instead of reciting; or expose independent control over the causes of a sound.

Beyond some threshold, the hidden state has to represent variables that are functionally equivalent to respiratory, laryngeal, articulatory, sensorimotor, autonomic, cognitive, and social bodily states. The model begins to operate as an embodied simulation in the causal sense, whether or not its developers describe it that way.

Three stages help clarify the argument:

Acoustic imitation: reproduce a waveform that resembles observed speech.
Embodied generativity: maintain a hidden organism-like state whose changes cause the waveform.
Embodied agency: perceive, predict, regulate, and adapt that state through interaction.

The progression is less like adding an “emotion” setting to speech and more like following the causal chain backward:

\text{waveform} \leftarrow \text{vocal apparatus} \leftarrow \text{whole-body physiology} \leftarrow \text{sensorimotor control} \leftarrow \text{cognitive, affective, and social state}

Where the shortcuts start to fail

Threshold	What begins failing without embodied state
1. Intelligibility	Phonemes, formants, and consonant timing
2. Local naturalness	Coarticulation, voice quality, and microprosody
3. Utterance coherence	Breath groups, fatigue, and phrase-scale pitch and intensity
4. Expressive realism	Smiling, tension, laughter, hesitation, and mixed emotion
5. Speaker persistence	Stable anatomy, habits, health, age, and vocal limits
6. Interactive realism	Listening while preparing, turn-taking, entrainment, and accommodation
7. Counterfactual control	”Say it while smiling nervously, whispering, walking uphill, and trying not to laugh”
8. Causal robustness	Novel combinations, pathology, perturbation, recovery, and adaptation

These thresholds are not a single ladder every system must climb in order. They mark places where a surface model is increasingly pressured to reconstruct the system that generated the surface.

I. Respiration: the power supply

#	What is heard	Speech or clinical term	What the model increasingly needs
1	Where inhalations occur	Speech breathing; breath-group planning	Lung volume, intended utterance length, syntactic plan, and available expiratory reserve
2	Phrase length	Utterance per breath; maximum phonation time	Vital-capacity-like limits, airflow consumption, and upcoming linguistic material
3	Loudness across a phrase	Subglottal pressure; SPL contour	Respiratory muscle recruitment and pressure decay
4	Pitch drifting as air runs down	Lung-volume effect on F0	Coupling among recoil pressure, laryngeal posture, and vocal-fold tension
5	Audible inhalation	Inspiratory noise; oral or nasal inspiration	Glottal opening, nasal versus oral route, flow rate, and urgency
6	Breathlessness	Dyspnea-like phonation; reduced breath support	Elevated respiratory drive, shortened phrases, rapid recovery breaths, and altered effort
7	Speech during exertion	Respiratory-phonatory competition	Locomotor demand, metabolic load, and the allocation of expiration between speech and gas exchange
8	Phrase-final weakening	Declination; terminal intensity fall; breathy offset	Falling pressure, laryngeal compensation, and the choice to trail off or replenish
9	Forceful projection	Vocal effort; respiratory drive	Abdominal and thoracic recruitment, subglottal pressure, and glottal resistance
10	Whisper duration and instability	Whisper airflow; turbulent excitation	High airflow consumption without periodic vocal-fold vibration
11	Sighs and voiced sighs	Paralinguistic expiration	Deep inspiration, passive recoil, and gradual laryngeal engagement
12	Coughs, throat-clears, and suppressed coughs	Protective laryngeal behaviours	Compression, glottal closure, explosive release, and competing social inhibition
13	Laugh-breath cycles	Egressive laughter pulses	Rapid intercostal contractions, expiratory bursts, phonated or unphonated pulses, and replenishing inhalations
14	Crying speech	Respiratory interruption; sobbing	Irregular inhalatory spasms, glottal constriction, and disrupted phrase planning
15	Breath timing in dialogue	Pre-turn inspiration; conversational breathing	Prediction of a turn opportunity, response preparation, and willingness to claim the floor

Speech breathing changes with utterance length and loudness. Conversational inhalation is tied to response planning and turn management. Laughter has its own respiratory organization; it is not an indefinitely repeatable string of ha tokens. [1]

II. The larynx: the voice source

#	What is heard	Relevant term	Required hidden variables
16	Fundamental frequency	F0; vocal-fold oscillation rate	Vocal-fold length, mass, longitudinal stiffness, and CT/TA muscle balance
17	Intensity	SPL; phonatory power	Subglottal pressure, glottal resistance, and closure behaviour
18	Breathy voice	Incomplete glottal closure; high H1-H2; spectral tilt	Resting glottal gap, adduction, tissue geometry, and airflow leakage
19	Pressed voice	Hyperadduction; low H1-H2; high contact quotient	Medial compression, intrinsic laryngeal activation, and pressure
20	Creak or vocal fry	Irregular phonation; pulse register; high jitter	Low F0, slack or short vocal folds, irregular closure timing, and low airflow
21	Falsetto or head-register quality	Register transition; thin-fold vibration	Reduced vibrating mass, altered TA/CT balance, and incomplete closure
22	Chest-register quality	Thicker medial surface; stronger harmonic excitation	Greater effective fold thickness and closure duration
23	Register breaks	Bifurcation; mode transition	Nonlinear tissue dynamics and instability across pressure-tension combinations
24	Smooth onset	Coordinated onset	Temporal alignment between airflow and vocal-fold adduction
25	Aspirate onset	Positive voice-onset interval	Airflow preceding complete laryngeal engagement
26	Hard or glottal onset	Glottal attack	Strong prephonatory adduction followed by abrupt pressure-driven oscillation
27	Whisper-to-voice transition	Phonation threshold pressure	Pressure, fold approximation, stiffness, and nonlinear onset conditions
28	Jitter	Cycle-to-cycle F0 perturbation	Oscillatory instability, motor noise, and tissue asymmetry
29	Shimmer	Cycle-to-cycle amplitude perturbation	Pressure variation, closure variation, and source instability
30	Harmonics-to-noise ratio	HNR; cepstral peak prominence	Periodicity, turbulent leakage, and regularity of tissue vibration
31	Diplophonia or subharmonics	Period doubling; nonlinear phonation	Asymmetric or multimodal fold oscillation
32	Vocal tremor	Rhythmic F0 or amplitude modulation	Oscillatory motor commands across respiratory and laryngeal subsystems
33	Strain	Muscle-tension-dysphonia-like quality	Excess intrinsic and extrinsic laryngeal tension, elevated laryngeal posture, and restricted vibration
34	A momentary voice crack under emotion	Phonatory instability	Rapid autonomic and motor changes that push the system through a nonlinear boundary
35	Voicing distinctions	Laryngeal timing; VOT; aspiration	Precise coordination of oral release, glottal opening, and onset of oscillation
36	Tonal-language pitch	Lexical tone production	Deliberate, syllable-aligned laryngeal tension trajectories rather than generic sentence melody

Fold stiffness, thickness, glottal opening, and subglottal pressure all affect the resulting acoustics. They also interact nonlinearly. A generator with independent controls for pitch, breathiness, and loudness will eventually produce impossible combinations unless those controls share an underlying biomechanics. [2]

III. Source-filter coupling: where modular controls break

#	What is heard	Term	What must be simulated
37	Vowel-dependent changes in voice quality	Source-filter interaction	Acoustic feedback from the vocal tract into glottal flow and fold vibration
38	Pitch instability near a formant	Nonlinear source-tract coupling	F0-resonance proximity, epilaryngeal geometry, and acoustic loading
39	Loudness efficiency	Inertive reactance; impedance matching	A vocal-tract configuration that assists or opposes glottal oscillation
40	Register behaviour changing with vowels	Filter-dependent phonation	Joint simulation of larynx and tract instead of independent pitch and timbre sliders
41	Singing resonance strategies	Formant tuning	Deliberate tract reshaping around source harmonics
42	Roughness caused by tract configuration	Source destabilization	Pressure feedback from supraglottal and subglottal cavities

The classical linear source-filter decomposition remains useful, but living anatomy is less tidy. Aerodynamic and acoustic coupling allow tongue position, laryngeal height, epilaryngeal geometry, and tract resonances to change the source itself. [3]

IV. Supralaryngeal articulation: the moving acoustic cavity

#	What is heard	Phonetic term	Required simulation
43	Vowel identity	F1, F2, F3; vowel space	Tongue-body shape, jaw aperture, lip posture, and pharyngeal geometry
44	Speaker-consistent vowels	Vocal-tract-length normalization	Stable anatomy plus habitual articulatory targets
45	Lip rounding	Labialization; lowered formants	Lip protrusion, aperture, and effective tract length
46	Smiling speech	Lip spreading; apparent tract shortening	Lip-corner displacement, cheek tension, and altered cavity geometry
47	Bilabials	Oral closure; burst release	Upper and lower lip contact, intraoral pressure, and release timing
48	Alveolars	Tongue-tip constriction	Tongue-tip or blade placement and release
49	Velars	Dorsal constriction; velar pinch	Tongue dorsum, palate geometry, and vowel-dependent contact location
50	Sibilants	Spectral centre of gravity; spectral moments	Jet formation, a grooved tongue, obstacle position, and dental geometry
51	Stops	Closure, burst, aspiration, VOT	Sealing, pressure accumulation, release, and laryngeal coordination
52	Fricatives	Turbulent source; constriction degree	Airflow through a narrow constriction and the geometry of a downstream obstacle
53	Affricates	Stop-fricative coordination	Controlled transition from complete closure to sustained constriction
54	Rhotic variation	Alveolar tap, trill, approximant, or uvular R	Distinct tongue or uvular gestures, airflow, and, for trills, self-sustained tissue vibration
55	Laterals	Lateral channels; antiformants	Central tongue contact with side-channel airflow
56	Nasals	Velopharyngeal opening; nasal formants and antiformants	Velum position, nasal-cavity geometry, and oral side branches
57	Nasalized vowels	Coupled oral-nasal resonance	Continuously graded velopharyngeal port opening
58	Hypernasality	Velopharyngeal insufficiency	Persistent or mistimed oral-nasal coupling
59	Nasal emission	Turbulent nasal airflow	Pressure, port area, oral obstruction, and nasal resistance
60	Clicks and ejectives	Non-pulmonic airstream mechanisms	Multiple closures, cavity-pressure changes, and glottalic or lingual airflow generation
61	Fine consonant transitions	Formant transitions	Continuous articulator trajectories rather than phoneme-wise concatenation
62	Reduced casual speech	Lenition; undershoot; assimilation	Effort-sensitive trajectories and incomplete attainment of ideal targets
63	Clear speech	Hyperarticulation; expanded vowel space	Increased movement amplitude and duration, guided by listener-oriented motor goals
64	Mumbling	Hypoarticulation	Reduced jaw excursion, compressed vowel space, and lower articulatory effort

Measured vocal-tract geometry can be converted into area functions that predict resonance. Even small changes in tract length affect perceived body size and speaker category; smiling changes acoustics partly because spreading the lips shortens the tract. [4]

V. Coordination and coarticulation: speech is not a row of sounds

#	What is heard	Term	What must be simulated
65	A consonant changing with its vowel context	Anticipatory or carry-over coarticulation	Overlapping articulatory gestures and context-sensitive targets
66	Natural syllable transitions	Gestural phasing	Relative timing of tongue, jaw, lips, velum, and larynx
67	Assimilation across word boundaries	Connected-speech processes	Planning beyond the current phoneme or word
68	Segment-duration changes	Compensatory timing	Redistribution of movement time under rate, emphasis, and complexity constraints
69	Consonant-induced pitch perturbations	Microprosody	Mechanical and neural coupling between laryngeal gestures and segment type
70	Stress changing articulation	Prosodic strengthening	Larger, longer, or more precise gestures in prominent positions
71	Phrase-edge articulation	Domain-initial strengthening; final lengthening	Hierarchical prosodic planning acting on motor trajectories
72	Speech-rate effects	Spatial and temporal reduction	Biomechanical velocity and acceleration limits, plus target undershoot
73	The tongue compensating for a blocked jaw	Motor equivalence	Task-level acoustic goals and redundant articulatory solutions
74	Stable speech despite perturbation	Sensorimotor compensation	Online detection of bodily error and redistribution across articulators
75	Speaker-specific rhythm	Idiolectal motor timing	Persistent learned coordination patterns rather than a global tempo control
76	Realistic speech errors	Phonological versus motor errors	Multiple planning levels whose failures produce different kinds of output

Articulatory synthesis improves when consonants are generated as context-dependent vocal-tract configurations. Human speakers also preserve acoustic goals by compensating with unperturbed articulators when, for example, the jaw is unexpectedly blocked. [5]

VI. Sensorimotor and neurological control

#	What is heard	Neurological term	Hidden process required
77	Fluent learned sequences	Feedforward motor commands	Stored mappings from phonological targets to coordinated gestures
78	Pitch correction after altered self-hearing	Auditory feedback control; pitch-shift reflex	Predicted auditory target, heard output, and error correction
79	Correction of jaw or tongue position	Somatosensory feedback control	Proprioceptive and tactile estimates of articulator state
80	Rapid self-correction	Efference copy; state estimation	Prediction of the sensory result before delayed feedback arrives
81	Stable output in noise	Lombard effect	Perceived noise, intelligibility estimates, and compensatory intensity or articulation
82	Speech adapted to room acoustics	Sidetone; external auditory feedback	Hearing reflected voice and adjusting effort
83	Sequencing syllables	Speech motor programming; GODIVA-like planning	Competitive selection and buffering of forthcoming motor chunks
84	Timing precision	Cerebellar prediction	Millisecond-scale coordination and error-based adaptation
85	Initiation and scaling	Basal-ganglia motor regulation	Movement onset, amplitude, automaticity, and vigour
86	Speech under divided attention	Dual-task interference	Shared executive and motor resources
87	Hesitation during lexical search	Incremental formulation	A partial linguistic plan interacting with a ready but delayed motor plan
88	Apraxic inconsistency	Apraxia of speech	Impaired planning or programming despite available muscles
89	Dysarthric weakness or incoordination	Dysarthria	Impairment distributed across respiration, phonation, resonance, articulation, and prosody
90	Stuttering-like disruption	Initiation or sequence instability	Competition, timing, and feedback dynamics rather than inserted textual repetitions
91	Motor adaptation over a conversation	Sensorimotor learning	Persistent updates to predicted body-sound mappings
92	Reflexive versus voluntary vocalizations	Dual laryngeal control systems	Distinct control pathways for speech, laughter, crying, and airway protection

Speech production is a distributed control problem. Neural models such as DIVA distinguish feedforward commands, auditory targets, and somatosensory targets. Perturbation studies show speakers compensating through bodily feedback, while altered auditory feedback changes voice production. [6]

VII. Whole-body geometry and condition

#	What is heard	Relevant term	What must be simulated
93	Apparent body size	Vocal-tract length; formant scaling	Stable head, pharynx, and tract geometry
94	Apparent age	Developmental vocal-tract morphology	Growth-dependent tract proportions and vocal-fold development
95	Sex-linked morphology cues	F0-VTL interaction	Fold dimensions and tract geometry, not pitch alone
96	Posture-dependent resonance	Postural phonation; tract deformation	Head, neck, jaw, spine, and laryngeal position
97	Voice while lying down	Supine versus upright tract acoustics	Gravity-sensitive soft tissue and airway geometry
98	Head turning during speech	Orientation-dependent radiation	Mouth direction, neck configuration, and changing microphone or listener geometry
99	Speech while smiling	Orofacial expression	Coordinated lips, cheeks, jaw, and sometimes laryngeal or respiratory changes
100	Speech while gesturing	Co-speech motor coupling	Shared timing and effort across hand, torso, jaw, and tongue movements
101	Walking or dancing speech	Locomotor-respiratory entrainment	Movement rhythm, impacts, exertion, and respiratory competition
102	Eating or dry-mouth speech	Oral lubrication; altered articulation	Saliva, mucosal contact, swallowing interruptions, and oral obstruction
103	Morning voice	Tissue state; mucus; edema-like change	Time-varying fold mass, hydration, and secretion state
104	Vocal fatigue	Laryngeal muscle or tissue fatigue	Accumulated loading, reduced control, altered effort, and recovery
105	Illness or congestion	Upper-airway obstruction; dysphonia	Nasal resistance, secretions, inflammation, and altered resonance
106	Pain-modified speech	Protective guarding	Restricted movement, breath-holding, and effort avoidance
107	Medication or intoxication effects	Sensorimotor alteration	Arousal, timing, muscle tone, coordination, and salivation changes
108	Stable bodily limits	Speaker-specific range profile	Physiological boundaries on F0, SPL, duration, and voice quality

Posture changes vocal-tract acoustics; body morphology affects tract length and formants; vocal effort changes with room acoustics and accumulated fatigue. Co-speech gesture can even alter tongue and jaw displacement, which means “voice only” is not always motorically separable from the visible body. [7]

VIII. Autonomic, affective, and interoceptive state

#	What is heard	Term	Required state
109	Stress-related pitch change	Autonomic arousal; F0 elevation or variability	Sympathetic activation interacting with laryngeal tension and respiratory drive
110	A tight, effortful voice	Perilaryngeal tension; hyperfunction	Extrinsic and intrinsic muscle recruitment
111	Quivering fear	Tremulous phonation	Unstable motor drive, respiratory irregularity, and heightened arousal
112	Anger	High-arousal prosody	Increased pressure, intensity, rate, articulation, and laryngeal activation
113	Sadness	Low activation; reduced prosodic range	Lower movement vigour, intensity, and pitch range, with slower timing
114	Excitement	Positive high arousal	Elevated respiratory drive and pitch variation without anger’s constrictive configuration
115	Calm warmth	Low arousal; positive valence	Low tension with controlled breath, smooth onset, and flexible resonance
116	Nervous smiling	Affect-expression conflict	Simultaneous lip spreading, laryngeal tension, uncertain timing, and inhibited laughter
117	Trying not to laugh	Suppression; leakage	Competing voluntary inhibition and involuntary respiratory or laryngeal laughter patterns
118	Trying not to cry	Affective suppression	Glottal constriction, disrupted breathing, swallowing, and pitch instability
119	Confidence versus uncertainty	Epistemic prosody	Commitment state affecting tempo, terminal contours, intensity, and hesitation
120	Cognitive load	Voice stress; planning load	Executive demand, autonomic response, and disrupted motor precision
121	Emotional mixtures	Blended affective prosody	Multiple partly independent systems rather than a single emotion label
122	Emotion changing during a sentence	Dynamic affect trajectory	Continuous latent state evolving faster or slower than linguistic structure
123	Genuine versus posed laughter	Spontaneous or volitional laughter acoustics	Different respiratory, laryngeal, and social-control pathways
124	Contagious laughter and vocal entrainment	Social-affective coupling	Perception of another person changing the speaker’s autonomic and motor state

Stress is associated with changes in laryngeal muscle activity and acoustic output, yet no single acoustic cue uniquely identifies an emotion. Faithful synthesis therefore needs a multidimensional physiological state rather than a lookup table that translates emotion=angry into “raise pitch and volume.” [8]

IX. Prosody, phonology, and linguistic embodiment

#	What is heard	Term	Embodied dependency
125	Lexical tone	Tone contour; tone sandhi	Laryngeal trajectories synchronized with syllable structure
126	Question versus statement	Boundary tone	Phrase-level motor plan and pragmatic intention
127	Which word is emphasized	Nuclear accent; prominence	Joint F0, duration, intensity, and articulatory strengthening
128	Contrastive focus	Focus prosody	A listener model and selection among alternative meanings
129	Stress-timed or syllable-timed rhythm	Rhythmic organization	Language-specific timing attractors imposed on articulatory gestures
130	Mora timing	Moraic phonology	Fine duration control spanning segments and syllables
131	Gemination	Contrastive consonant duration	Longer closure or constriction without merely slowing a recording
132	Phonemic vowel length	Quantity contrast	Duration integrated with stress, rhythm, and articulation
133	Accent and dialect	Sociophonetics; phonological grammar	Learned motor targets, timing patterns, and social-indexical control
134	Native-like rolled R	Language-specific rhotic gesture	The right articulator, airflow, timing, and phonological environment
135	Code-switching	Language-mode switching	Reconfiguration of phonetic targets, prosody, rhythm, and social stance
136	Aizuchi and backchannels	Continuers; response tokens	Timing, pitch, intensity, and relational stance, often with little lexical content
137	Sarcasm or irony	Pragmatic prosody	Divergence between literal semantic content and social intention
138	Quotation voice	Enactment; reported speech	A temporary shift in speaker model and bodily stance
139	Vocal stimming and playful sound	Nonlexical vocalization; paraspeech	Sensorimotor pleasure, repetition, rhythmic attractors, and expressive control not reducible to text
140	Whispered tone or stress	Prosody without periodic F0	Substitute cues such as duration, intensity, spectral shape, and airflow
141	Singing-to-speech transitions	Speech-song continuum	Shared but differently constrained respiratory, laryngeal, and articulatory systems

X. Interactional embodiment

#	What is heard	Interaction term	What must be simulated
142	Knowing when someone has finished	Turn-final prosody	Prediction from syntax, pitch, duration, gaze, context, and breath
143	Preparing before the other person stops	Anticipatory turn planning	Concurrent listening, response formulation, and pre-inspiratory preparation
144	Brief overlaps	Competitive or cooperative overlap	Urgency, affiliation, and predicted completion time
145	Holding the floor	Turn-holding cues	Continuing intonation, filled pauses, breath management, and pacing
146	Yielding the floor	Turn-yielding cues	Terminal contour, slowing, intensity reduction, and completion
147	Matching another person’s rhythm	Entrainment; accommodation	Online estimation and gradual adaptation of timing, F0, and intensity
148	Mirroring vocal fry or accent	Phonetic convergence	Social relation plus adjustable motor habits
149	Deliberately refusing to mirror	Divergence; stance marking	Metacontrol over accommodation
150	Speaking to a child, elder, or hard-of-hearing listener	Audience design; clear speech	A listener-specific model of intelligibility
151	Intimate versus public voice	Register; interpersonal distance	Breathiness, loudness, articulation, and social safety
152	Micro-signals of acceptance	Affiliative prosody	Timing, softness, tolerance of overlap, and backchannel selection
153	Detecting a joke before its words finish	Prosodic framing	Shared context and predictive social inference
154	Laughing together	Laughter entrainment	Reciprocal timing, breath cycles, and sensitivity to the other person’s state
155	Speaking in a noisy room	Lombard adaptation	Environmental sensing, listener distance, and vocal effort
156	Speaking in a reverberant room	Room-voice adaptation	External auditory feedback and changed effort
157	A different voice for a “formal” task	Style shifting; read versus spontaneous speech	Social role, monitoring level, and learned performance posture
158	Continuity through interruptions	Conversational state	Persistent bodily, emotional, and discourse state across turns

Human turn transitions are extremely fast relative to the time required to plan speech. Listening and embodied response preparation must overlap. Breathing participates in the organization of turns, while face-to-face conversation also recruits gaze and gesture. [9]

XI. Pathology as a causal test

A model can produce generic healthy speech from correlations alone. Pathology is a stronger test of whether it has learned the causal subsystems.

#	Audible presentation	Underlying subsystem that must be represented
159	Flaccid dysarthria	Weakness, reduced tone, breath support, and incomplete closure
160	Spastic dysarthria	Hypertonia, strained phonation, and slow, restricted movement
161	Ataxic dysarthria	Variability in timing, force, and coordination
162	Hypokinetic dysarthria	Reduced movement scaling, monopitch, monoloudness, and accelerated bursts
163	Hyperkinetic dysarthria	Involuntary movement entering phonation and articulation
164	Apraxia of speech	Impaired motor planning, segmentation, and inconsistent errors
165	Muscle tension dysphonia	Maladaptive extrinsic or intrinsic laryngeal recruitment
166	Vocal-fold paresis	Asymmetric closure and vibration
167	Nodules or lesions	Changed mass, stiffness, closure, and periodicity
168	Velopharyngeal insufficiency	Abnormal nasal coupling and pressure loss
169	Cleft-related speech	Structural resonance and pressure constraints plus compensatory articulation
170	Parkinsonian speech	Basal-ganglia-linked changes in scaling and initiation
171	Cerebellar disease	Timing and coordination deficits
172	ALS-related speech	Progressive subsystem weakness with changing compensations
173	Hearing-loss-related speech	Altered long-term auditory calibration
174	Post-laryngectomy speech	A different sound source and different source-tract coupling
175	Tracheostomy or respiratory disease	Altered airflow, pressure, and phrase limits
176	Stuttering	State-dependent initiation, timing, tension, and adaptation
177	Functional voice disorder	A learned control policy and context-dependent tension without a simple structural lesion
178	Recovery or therapy effects	Longitudinal adaptation of motor programs, compensation, and effort

The traditional speech-pathology analysis of respiration, phonation, resonance, articulation, and prosody is almost a ready-made specification for the minimum modular body model needed by clinically faithful synthesis. Dysarthria can reflect impairment across several of these systems, which makes a single “dysarthria style” embedding causally inadequate. [10]

XII. What emerges above the body parts

Once the lower layers are modelled, several larger latent structures become necessary.

179. A persistent body schema

The model must remember tract length, pitch range, lung capacity, habitual posture, speech rate, asymmetries, and current fatigue. Otherwise the apparent body silently changes between sentences.

180. A feasible-action manifold

Not every combination of pitch, intensity, breathiness, speed, and articulatory precision is physically reachable. A causal model needs a representation of possible actions, including the cost of moving between states.

181. Effort

Two acoustically similar outputs can demand different levels of muscular and cognitive effort. That difference affects what comes next: fatigue, breath timing, errors, recovery, and willingness to continue.

182. Interoceptive state

The generator needs functionally equivalent estimates of air hunger, strain, dryness, fatigue, tension, and instability. They need not be felt qualities. They do need to be available to control.

183. Forward prediction

Before producing a sound, the system must predict what a planned motor state will sound like.

184. Self-perception

It must register its generated output as an event caused by itself, compare that output with its target, and distinguish self-generated sound from environmental sound.

185. Error attribution

When output differs from prediction, the system must infer whether the cause lies in the larynx, articulation, breath, room, microphone, noise, listener interruption, or the plan itself.

186. Compensation

It must find another bodily configuration that preserves the communicative target when the preferred configuration is unavailable.

187. History and hysteresis

The same command can produce a different output depending on the preceding state. A tired, dry, recently laughing, or highly tensed system does not reset between tokens.

188. Multiple timescales

Voice contains interacting dynamics that operate over:

milliseconds: vocal-fold cycles and release bursts;
tens of milliseconds: articulatory gestures;
hundreds of milliseconds: syllables and pitch accents;
seconds: breath groups and conversational turns;
minutes: entrainment and fatigue;
years: accent, pathology, age, and vocal training.

189. An interlocutor model

Prosody is partly an action on another nervous system. Loudness, clarity, timing, emphasis, tenderness, and hesitation depend on what the speaker model predicts the listener will hear, know, feel, and do.

190. A self-model

The voice must be generated as this speaker, with this body, under these conditions, attempting this act. Speaker identity cannot remain a static timbre vector once the system is expected to behave coherently outside its training distribution.

The embodied-simulation hypothesis

The hypothesis can now be stated precisely:

As vocal synthesis approaches causal, counterfactual, and interactive fidelity, the minimal sufficient model of voice expands from an acoustic distribution toward a generative model of an embodied, regulating, socially situated organism.

More formally, for every acoustically detectable bodily variable $B_i$ , a synthesizer expected to generalize across interventions on $B_i$ , interactions among such variables, and previously unseen combinations cannot rely indefinitely on surface correlations. It must encode a latent state $Z_i$ that preserves enough of the causal structure of $B_i$ to predict its downstream acoustic consequences.

The latent state does not have to resemble a photorealistic human body. If it preserves:

state,
dynamics,
constraints,
coupling,
counterfactual response,
adaptation,
memory, and
action possibilities,

then it is a body model by functional equivalence.

This does not by itself prove consciousness. It does undermine the simple version of the “stochastic parrot” picture in which high-fidelity vocal behaviour is understood as merely selecting plausible surface sequences. A model may begin by imitating recordings, but the target distribution was generated by organisms. At sufficiently high resolution, the regularities of the organism are embedded in the data, and reproducing them robustly requires reconstructing more and more of the generating system.

Speech is not text decorated with audio. It is audible physiology, controlled by a nervous system and situated inside an interaction.