tonalityprint/tonalityprint-v1

TonalityPrint Voice Dataset v1.0

A Contrast-Structured Voice Dataset for Exploring Functional Tonal Intent, Ambivalence, and Inference-Time Prosodic Alignment

---

📥 DOWNLOAD DATASET FILES

⚠️ This Hugging Face repository contains DOCUMENTATION ONLY.

Download audio and annotation files from Zenodo (official source):
https://doi.org/10.5281/zenodo.17913895

Why Zenodo?

• ✅ Official DOI for academic citations
• ✅ Permanent archival storage
• ✅ Download statistics for grant reporting
• ✅ Academic credibility

Quick Download: DATACARD.zip (42.9 MB)

---

Overview

TonalityPrint is a specialized single-speaker speech corpus designed to enable exploration of fine-tuning functional tonal intents in voice AI systems. Unlike emotion recognition datasets, TonalityPrint annotates functional tonal intents (what speakers do with tone), not just what they feel.

Key Features:

• 144 high-fidelity WAV files (48kHz, 16-bit, mono, unprocessed)
• 18 unique utterances across 8 parallel prosodic states
• 5 functional tonal intents: Trust, Attention, Reciprocity, Empathy Resonance, Cognitive Energy
• Continuous intensity indices (0-100 scale) for each intent
• Ambivalence annotation (perceptual entropy cross-intent feature)
• 100% authentic human voice with explicit consent
• Single-speaker design eliminates speaker variability for controlled analysis

What This Dataset Is:

• A precision-tuning resource for prosodic AI alignment research
• A controlled substrate for investigating functional tonal intent
• An experimental framework for ambivalence-aware dialogue systems
• A hypothesis-generating tool for human-AI voice calibration

What This Dataset Is Not:

• A general-purpose emotion recognition training corpus
• A multi-speaker dataset for population-level generalization
• A substitute for large-scale speech datasets
• A validated benchmark for production systems

---

Dataset Composition

Structure

From Zenodo Download:

DATACARD/
├── audio/                              # 144 WAV files
├── annotations/
│   ├── json/                          # 144 JSON files
│   ├── csv/                           # 144 CSV files
│   └── ALL_TONALITY_DATA_COMBINED.csv # Combined dataset
└── documentation/                      # Technical references

Audio Specifications

Specification	Value
Format	WAV (uncompressed PCM)
Sample Rate	48,000 Hz (48kHz)
Bit Depth	16-bit
Channels	Mono (1 channel)
Duration per File	3-6 seconds
Total Duration	~11 minutes 5 seconds
Processing	None (raw, unprocessed)
Total Files	144 audio samples

Fixed-Phrase Octet Design

The dataset uses a Fixed-Phrase Octet structure: 18 utterances × 8 parallel prosodic states.

Each utterance is recorded in:

1. Baseline/Neutral (control sample)
2. Trust (Trus) - conveying reliability and credibility
3. Attention (Att) - directing focus and engagement
4. Reciprocity (Reci) - expressing mutual exchange
5. Empathy Resonance (Emre) - demonstrating empathetic connection
6. Cognitive Energy (Cogen) - showing mental engagement
7. Sub-modified variants (e.g., Trust + Calm)
8. Ambivalence variants (optional cross-intent complexity)

This design enables:

• Differential Latent Analysis (DLA): Isolate prosodic features while holding lexical content constant
• Contrastive learning: Compare prosodic variations across identical text
• Intent vector extraction: Model functional intent as steerable features

---

Controlled Semantic Design

Functional Tonal Intents (Not Emotions)

TonalityPrint distinguishes between functional intent and affective state:

Functional Intent	What It Does	Not The Same As
Trust	Establishes credibility, reliability	"Happiness" or "Confidence"
Attention	Directs focus, maintains engagement	"Excitement" or "Urgency"
Reciprocity	Invites response, balances exchange	"Friendliness" or "Agreement"
Empathy Resonance	Attunes to listener state	"Sympathy" or "Sadness"
Cognitive Energy	Signals mental activation	"Enthusiasm" or "Anxiety"

Why This Matters:

• Traditional emotion datasets label what speakers feel
• TonalityPrint annotates what speakers do with their voice
• This functional framing aligns with conversational AI goals

Ambivalence as Feature (Not Noise)

Unlike traditional datasets that discard mixed signals as annotation errors, TonalityPrint systematically annotates ambivalence (ambivalex) as:

• A perceptual entropy transitional state
• A cross-intent feature where competing tonal cues co-occur
• An essential signal for real-world inference-time alignment

---

How to Use

Download from Zenodo

# 1. Visit Zenodo
https://doi.org/10.5281/zenodo.17913895

# 2. Download DATACARD.zip (42.9 MB)

# 3. Extract files
unzip DATACARD.zip

Load Annotations

import pandas as pd

# Load combined CSV
df = pd.read_csv('DATACARD/annotations/ALL_TONALITY_DATA_COMBINED.csv')

print(f"Total samples: {len(df)}")
print(f"Columns: {df.columns.tolist()}")

# Filter by intention
trust_samples = df[df['Primary_Intention'] == 'Trust']
ambivalent_samples = df[df['Ambivalex'] == 'ambivalex']

Load Audio Files

import librosa

# Load audio file
audio_path = 'DATACARD/audio/TPV1_B1_UTT1_S_Att_SP-Ronda.wav'
audio, sr = librosa.load(audio_path, sr=48000, mono=True)

print(f"Sample rate: {sr} Hz")
print(f"Duration: {len(audio)/sr:.2f} seconds")

# Extract features
mfcc = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=13)

Explore Tonality Indices

# Compare Trust scores across utterances
trust_scores = df.groupby('Utterance_Number')['Trust_Index'].mean()

# Analyze Cognitive Energy bias
ce_by_intent = df.groupby('Primary_Intention')['Cognitive_Energy_Index'].describe()

---

Annotation Methodology

Expert Practitioner Annotation

Annotator: Ronda Polhill (speaker and dataset creator)
Method: Expert perceptual assessment combined with acoustic analysis
Expertise: 8,873+ high-stakes customer interactions (observational context, not causal proof)

Continuous Indices (0-100 Scale)

Each utterance includes five tonality indices:

Index	Abbreviation	Interpretation
Trust	TR	0-30: Low/Minimal, 31-60: Moderate, 61-85: High, 86-100: Very High
Attention	AT	Perceptual score of attentional focus
Reciprocity	RE	Perceptual score of collaborative tone
Empathy Resonance	ER	Perceptual score of empathetic attunement
Cognitive Energy	CE	Perceptual score of mental activation

Important: These are annotator perceptual scores, not empirically validated scales.

Quality Control

• Proprietary heuristic audit: ~80+% acoustic-intent alignment verified
• Re-recording rate: ~18.05% of corpus re-recorded for consistency
• Known bias: Cognitive Energy shows systematic elevation (documented and retained)

---

Intended Use

Primary Applications

1. Inference-Time Prosodic Alignment

   • Fine-tuning reasoning-based voice models
   • Aligning model confidence with vocal uncertainty
   • Calibrating trust signals in AI responses

2. Differential Latent Analysis

   • Extracting tonal intent vectors (analogous to LLM activation steering)
   • Contrastive learning with fixed lexical content
   • Isolating prosodic features from semantic content

3. Ambivalence-Aware Systems

   • Training dialogue systems to detect mixed signals
   • Modeling uncertainty in safety-critical applications
   • Navigating tonal complexity in nuanced interactions

4. Style-Conditioned Synthesis

   • Controlling prosodic style in TTS systems
   • Evaluating voice quality metrics
   • Transfer learning for expressive speech

5. Human-AI Voice Calibration

   • Investigating "AI-adjacent yet trusted" vocal profiles
   • Studying uncanny valley effects in voice
   • Exploring voice-appearance synchrony in embodied AI

---

Known Biases and Limitations

Single-Speaker Constraint

• All 144 files from same speaker (Ronda Polhill)
• Findings may not generalize across:
  • Genders, ages, accents, cultures, languages
• Multi-speaker validation required for broader applicability

Cognitive Energy Systematic Bias

Known Issue: Cognitive Energy Index shows systematic elevation across corpus.

Possible Causes:

• Speaker's natural ecological style (high-energy delivery)
• Lexical content effects
• Practitioner annotation bias

Resolution: Intentionally retained for transparency. Researchers should account for this bias in analyses.

Controlled Environment

• Professional studio recordings (not naturalistic)
• Scripted content (not spontaneous speech)
• May not reflect real-world acoustic conditions

Observational Context (Not Causal Proof)

The annotation methodology references 8,873+ customer interactions with observed correlations:

• ~35.85% average conversion rate (observational metric)
• 68 spontaneous reports of "AI-adjacent" voice quality with high trust ratings

Critical Caveat: These are observational correlations, not causal relationships. Multiple confounding variables present.

---

Ethical Considerations

Speaker Consent and Biometric Integrity

• 100% human recordings by author (Ronda Polhill)
• Explicit informed consent for recording, annotation, and public release
• No synthetic voices, clones, or generative AI audio
• Speaker demographics: Mid-life female, native English speaker

Prohibited Uses

Researchers are strictly prohibited from:

• Creating unauthorized voice clones of the speaker
• Generating deepfakes using this dataset
• Using recordings for deceptive purposes
• Violating CC BY-NC 4.0 license terms

---

Links

• 📊 Official Dataset (Zenodo): https://doi.org/10.5281/zenodo.17913895
• 📚 Documentation (GitHub): https://github.com/YOUR_USERNAME/TonalityPrint-v1
• 📄 White Paper: https://doi.org/10.5281/zenodo.17410581
• 🌐 Website: https://TonalityPrint.com
• 📧 Contact: ronda@TonalityPrint.com

---

Citation

BibTeX

@dataset{polhill_2026_tonalityprint,
  author       = {Polhill, Ronda},
  title        = {TonalityPrint: A Contrast-Structured Voice Dataset 
                  for Exploring Functional Tonal Intent, Ambivalence, 
                  and Inference-Time Prosodic Alignment v1.0},
  year         = 2026,
  publisher    = {Zenodo},
  version      = {1.0.0},
  doi          = {10.5281/zenodo.17913895},
  url          = {https://doi.org/10.5281/zenodo.17913895}
}

APA

Polhill, R. (2026). TonalityPrint: A Contrast-Structured Voice Dataset for Exploring Functional Tonal Intent, Ambivalence, and Inference-Time Prosodic Alignment v1.0 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.17913895

---

License

CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International)

• ✅ Academic and research use: FREE
• ✅ Proper attribution required
• ❌ Commercial use: Requires licensing

Commercial licensing: Contact ronda@TonalityPrint.com

---

Acknowledgments

This work emerges from independent practitioner-research conducted without institutional funding and is released for academic research use under CC BY-NC 4.0.

TonalityPrint aims to address a critical gap in voice AI training data by moving beyond discrete emotion recognition to capture functional tonal intent, including ambivalent prosodic signals as essential nuances for inference-time alignment.

---

Version: 1.0.0
Release Date: January 24, 2026
Last Updated: January 30, 2026
License: CC BY-NC 4.0
© 2026 Ronda Polhill