Prosodic Entrainment Optimizer

Vocal entrainment is the natural, subconscious phenomenon where people synchronize their speech patterns — pitch, pace, rhythm, and intensity — with their conversation partner. This synchronization is one of the most robustly documented proxies for rapport, trust, and cooperative intent in human communication, first formalized by Howard Giles' Communication Accommodation Theory (1973) and now computationally validated through deep learning (Nasir et al., IEEE TAFFC 2022).

The supplied Python helper is an offline, stateless heuristic over supplied numeric features. It returns a score and suggested TTS deltas; it does not capture audio, inject TTS settings, log adjustments, or measure caller rapport. Audio extraction, timing, consent, and any application of suggestions belong to a separate host integration. It is not a clinical, crisis-response, or financial decision tool.

Research-inspired prototype: the studies below motivate the design but do not validate this implementation or its outcomes.

Scientific Foundation

Paper	Venue	Year	Contribution to This Skill
Communication Accommodation Theory	Language Sciences, Elsevier	2023	Theoretical framework: convergence/divergence mechanisms, mediated communication model
Modeling Vocal Entrainment via Deep Unsupervised Learning	IEEE Transactions on Affective Computing	2022	Triplet-network entrainment distance → basis for cosine similarity aggregation
Context-Aware Computational Entrainment in Dyadic Conversations	arXiv	2022	Cross-subject attention model for real-time dyadic entrainment tracking
ISO/IEC 42001:2023 AI Management System	ISO	2023	Transparency, logging, and auditability requirements for adaptive AI systems

How It Works

A proposed host integration could place this helper between feature extraction and TTS. The shipped helper only computes suggestions:

Step 1: Baseline Calibration (first 5 seconds)

The host may set is_calibrating=True during a chosen warm-up window. This returns CALIBRATING and an identity directive; the helper itself has no clock, feature collection, or automatic five-second calibration.

Step 2: Feature Extraction (per 1-second window)

The host must supply three features; no audio extractor is included:

• Fundamental Frequency (F0, Hz): Perceived pitch — speaker identity and emotional state marker.
• Speech Rate (WPM): Cognitive load and urgency indicator.
• RMS Energy (normalized): Loudness and engagement level.

Step 3: Entrainment Score Computation

Cosine similarity is computed between the L2-normalized caller and agent prosodic feature vectors. entrainment_score ∈ [0.0, 1.0]:

• 1.0 = perfect prosodic synchrony
• 0.0 = complete divergence

Step 4: Status Classification & TTS Directive Generation

entrainment_score	Status	Action
>= 0.90	OPTIMAL	No directive — back off to avoid over-mirroring
0.75 – 0.89	TARGET_REACHED	Bounded directive may still be returned
< 0.75	LOW_ENTRAINMENT	Issue TTSDirective with bounded parameter deltas
First 5s	CALIBRATING	Identity directive — no adjustment

Step 5: Safety-Bounded TTS Directive

Parameter	Adjustment Logic	Safety Cap
pitch_shift_semitones	Proportional to F0 delta (Hz → semitones)	±3.0 semitones/window
rate_multiplier	Proportional to WPM ratio	[0.80, 1.20]
energy_scale	Proportional to RMS energy ratio	[0.70, 1.30]

The formula uses a fixed 0.05 fraction of the feature difference, subject to the caps above. This is not a 5% output cap or a time-based rate limit; the host controls invocation timing.

Mode Presets

Mode	Use Case	Behavior
DEFAULT	General inbound/outbound	Balanced bidirectional convergence
SALES	Outbound sales, lead qualification	Converge toward caller's energy to build rapport
SUPPORT	Synthetic downward-only demonstration	Never raises pitch, rate, or energy; no clinical de-escalation efficacy is established

Key Features

• Bounded suggestions: Parameter caps are enforced per invocation; perceptual effects have not been measured.
• Anti-over-mirroring: Backs off automatically at score >= 0.90 to prevent the "uncanny valley" of identical-sounding voices.
• Host-controlled calibration: is_calibrating=True suppresses adjustments.
• Fail-safe on zero/silence: Muted callers, zero-rate speech, and whispering are all handled without crashes or division-by-zero errors.
• SUPPORT mode: Only holds or lowers pitch, rate, and energy. It does not implement an upward matching phase or establish de-escalation efficacy.

Configuration Reference

Source constants and proposed host settings are listed below, not a runtime configuration API. The helper has no window-duration or calibration-duration parameter.

Parameter	Default	Range	Description
TARGET_THRESHOLD	0.75	0.60 – 0.85	Below this → issue TTSDirective
OPTIMAL_CEILING	0.90	0.80 – 0.95	Above this → back off (no directive)
MAX_PITCH_DELTA	3.0 semitones	1.0 – 5.0	Safety cap on pitch adjustment per window
MAX_RATE_DELTA	0.20 (±20%)	0.10 – 0.30	Safety cap on rate multiplier delta
MAX_ENERGY_DELTA	0.30 (±30%)	0.15 – 0.40	Safety cap on energy scale delta
step_factor	0.05	Source edit only	Fixed interpolation factor, not an output/time cap
window_duration_s	1.0	0.5 – 2.0	Feature extraction window length
calibration_duration_s	5.0	3.0 – 10.0	Baseline collection period

Expected Outcomes & Metrics

The following are design hypotheses or operating targets, not measured outcomes of this helper:

Metric	Expected Improvement	Notes
CSAT Score	Not measured	Requires a separate evaluation
Call Abandonment Rate	Not measured	No outcome improvement is established
First Call Resolution (FCR)	Not measured	No outcome improvement is established
Entrainment Score (avg call)	0.78 – 0.85	Target operating range
Directive latency	< 10ms	Synthesis parameter update time

Use Cases

Proposed research contexts only, not validated clinical, crisis-response, sales, or financial deployments:

• Outbound sales calls: Mirror the prospect's energy and cadence to build trust before pitching.
• Healthcare intake: Automatically slow pace and lower pitch to match an elderly or anxious caller, reducing cognitive load and improving information capture.
• Debt collection: Reduce confrontational dynamics by actively converging toward a calm, measured pace even when the caller is agitated.
• Mental health support lines: Gently mirror distressed caller's cadence (SUPPORT mode) while guiding toward slower, regulated rhythm via Match & Lead.
• High-volume IVR exit: Reduce caller frustration after a failed IVR interaction by rapidly entraining to their speech pattern when a live-agent-style AI picks up.

Limitations & Known Constraints

• Cosine similarity is angle-based: Features that point in the same direction in feature space can score high even with different magnitudes. This is expected behavior — the system is measuring style convergence, not energy matching. The energy scale directive handles amplitude alignment separately.
• TTS dependency: The TTSDirective output requires a TTS engine that accepts real-time prosodic parameter overrides (e.g., SSML <prosody> tags or equivalent API). Not all TTS providers support this.
• Non-verbal speakers: The host must detect missing/unreliable features and choose whether to keep calibration enabled. The helper does not infer this condition.
• Accent diversity: Threshold calibration must include diverse vocal profiles to prevent systematic bias against non-standard prosodic patterns.

Integration

[Caller Audio]
      |
[ASR + Prosodic Feature Extractor]
      |
[call-prosodic-entrainment-optimizer]  <-- this skill
      |
  TTSDirective { pitch=-1.2, rate=0.94, energy=0.91 }
      |
[TTS Synthesis Engine]  <-- applies SSML prosody overrides
      |
[Agent Voice → Caller]

References

See references/research-papers.md for research inspiration; citations are not implementation validation. See references/safety.md for safety and host responsibilities; no ISO conformity is established. See references/examples.md for end-to-end scenario walkthroughs.