Speech Craft Analytics (SCA) is a data-driven intelligence platform that measures management tone, stress, confidence, and communication quality in corporate earnings calls. SCA converts raw CEO/CFO audio into quantitative behavioral signals that forecast post-earnings stock returns and reveal insights that text-only sentiment cannot capture.

SCA focuses on how executives speak—not just what they say—using speech science, machine learning, and proprietary linguistic models. The result is a high-resolution dataset of acoustic, linguistic, and behavioral indicators designed for investment research, risk management, and corporate communication.

SCA exists to answer a simple but hard question:

Where traditional NLP stops at the transcript, SCA quantifies tone, vocal tension, pacing, confidence, nervousness, arousal, and cognitive load—the behavioral layer of corporate disclosure. These signals are shown in SCA’s white paper to be predictive of medium-horizon excess returns.

SCA produces interpretable, model-ready features in three families:

Speaker-baseline-normalized, non-linear features such as:

• Pitch and pitch dynamics
   
• Jitter and shimmer (micro-instabilities linked to stress)
   
• Harmonicity, energy, amplitude
   
• Speech rate, pause structure, breathing rate
   
• Spectral features, MFCCs, PCA-audio embeddings
   
• Voice Delivery Quality (VDQ)
   
• Voice Arousal and Voice Valence
   

• Sentiment polarity (including earnings-specific lexicons)
   
• Strong/weak modal verbs and hedging
   
• Disfluencies, filler words, and repetitions
   
• Honesty phrases and clarity markers
   
• Complexity indices (e.g., Gunning Fog)
   
• Pronoun usage (#I, #We), numerical emphasis
   
• Domain-specific word lists (Litigious, Negative, etc.)
   

Non-linear combinations of speaker-normalized audio and linguistic style:

• Audio (baseline-relative vocal delivery)
   
• Composite (~2/3 Audio, ~1/3 Linguistic)
   
• VDQ (Voice Delivery Quality)
   
• Assertiveness (core / mgmt)
   
• Balanced Delivery (core / mgmt)
   
• Nervousness (core / mgmt, inverse-ranked)
   
• Arousal_mgmt (inverse-ranked calmness)
   
• Valence_mgmt
   
• Emotion probabilities: p_confident, p_neutral, p_nervous
   
• Derived voice_arousal:
  voice_arousal = (p_nervous − p_confident + 1) / 2
   

These features are designed to be both human-interpretable and stable for quant models and LLMs.

For each earnings call, SCA provides event-study-ready outputs:

• Excess returns (stock – equal-weight benchmark) for H = 1, 5, 10, 15, 20, 25, 30
   
• Decile assignments (D1–D10) for each signal
   
• Cluster-robust t-statistics for D10 and D10–D1 portfolios
   
• Controls for:
   
  • Earnings surprise (%)
     
  • Reporter effect (companies reporting vs. not reporting on the same day)
     

These controls allow investors to separate the effects of fundamental news from the incremental information in tone.

Each spoken sentence includes:

• Timestamps and speaker IDs (CEO, CFO, other)
   
• 200+ acoustic and linguistic metrics
   
• Emotion probabilities (p_confident, p_neutral, p_nervous)
   
• Voice arousal & valence
   
• Disfluency and repetition markers
   
• Sentiment, modality, and honesty signals
   

This structure enables fine-grained research, LLM training, and custom signal engineering.

SCA’s white paper demonstrates consistent, economically meaningful alpha across Russell 3000 earnings calls (CEO/CFO Q&A, price > $10, 2020–2025).

• Long-only excess returns of ~40–70 bps over H=15–30 days
   
• Many horizons show t-stats between 2 and 3
   

(Balanced_mgmt, Arousal_mgmt, Nervous_mgmt)

• Low nervousness / low arousal (inverse-ranked D10) produce ~50–60 bps excess return over 3–6 weeks
   
• Effects persist even after controlling for earnings surprise and reporter effect, confirming incremental signal value
   

• Clean, steadily rising alpha curve: 30–40 bps by H=20–30
   
• When adjusted for EPS surprise, VDQ becomes a short-horizon signal (H=1), indicating rapid market incorporation of the non-fundamental tone component
   

While transcript sentiment is predictive, Audio, Composite, VDQ, and management-trained nervousness/arousal factors often show stronger performance and add incremental explanatory power when combined with sentiment and surprise.

Executives often use neutral or reassuring language, while their vocal delivery signals tension, uncertainty, or over-confidence—often before fundamentals reveal the issue.

Top-decile SCA behavioral signals consistently deliver 40–60 bps of long-only excess return over 1–6 week horizons, with significant t-stats, even after controlling for earnings surprise and reporting effects.

SCA signals are structured for easy integration into:

• Factor models
   
• Stock-selection frameworks
   
• LLM-powered research tools
   
• RAG systems needing clean metadata
   

SCA adds the behavioral layer that text alone cannot capture