After scanning a voicemail with VoiceGuardAI, you'll receive a detailed analysis report. Understanding these results is crucial for making informed decisions about potential voice scams. This guide explains how to interpret the scan results, understand confidence scores, and recognize the acoustic features our AI analyzes.
Scan Result Classifications
VoiceGuardAI classifies voicemails into three main categories:
Likely Human
Our AI has determined that the voice in this voicemail is most likely from a real human speaker.
What This Means:
- The voicemail exhibits natural speech patterns consistent with human speech
- No significant markers of synthetic voice generation were detected
- The audio contains expected variations in pitch, rhythm, and other natural speech characteristics
Likely AI-Generated
Our AI has detected characteristics consistent with synthetic or manipulated audio in this voicemail.
What This Means:
- The voicemail contains patterns typically found in AI-generated or manipulated speech
- Unnatural acoustic features were detected that are uncommon in human speech
- The audio may have been created using voice cloning, text-to-speech, or other synthetic voice technologies
Uncertain
Our AI has detected some suspicious patterns, but not enough to make a definitive determination.
What This Means:
- The voicemail contains some characteristics of both human and synthetic speech
- The audio quality may be too poor for a definitive analysis
- The sample may be too short for reliable detection
- The voice may have been partially manipulated or processed in ways that make classification difficult
Understanding Confidence Scores
Each scan result includes a confidence score, which indicates how certain our AI is about its classification. Here's how to interpret these scores:
| Confidence Range | Interpretation | Recommended Action |
|---|---|---|
| 90-100% | Very high confidence in the result | You can generally rely on this classification |
| 75-89% | High confidence, but some uncertainty | Consider the classification reliable, but use additional verification if the content is suspicious |
| 60-74% | Moderate confidence with significant uncertainty | Use caution and additional verification methods |
| Below 60% | Low confidence, highly uncertain | Do not rely on this classification alone; use multiple verification methods |
Acoustic Features Analyzed
VoiceGuardAI analyzes over 500 acoustic features to determine if a voice is human or AI-generated. While we don't disclose the exact algorithms used (to prevent circumvention), here are some of the general categories of features we analyze:
Prosodic Features
These relate to the rhythm, stress, and intonation of speech:
- Pitch variations: Natural human speech has micro-variations in pitch that are difficult for AI to perfectly replicate
- Speech rhythm: The timing between words, syllables, and pauses often differs between human and synthetic speech
- Stress patterns: How emphasis is placed on different syllables and words
Spectral Features
These relate to the frequency distribution of the audio:
- Harmonic structure: The pattern of harmonics in the voice
- Formant transitions: How the resonant frequencies of the vocal tract change during speech
- Spectral balance: The distribution of energy across different frequency ranges
Temporal Features
These relate to timing aspects of the speech:
- Voice onset time: The timing between the release of a consonant and the beginning of vocal fold vibration
- Articulation rate: How quickly sounds are produced
- Pause patterns: The duration and placement of pauses in speech
Voice Quality Features
These relate to the quality and characteristics of the voice:
- Breathiness: The amount of audible breath in the voice
- Jitter and shimmer: Cycle-to-cycle variations in frequency and amplitude
- Vocal fry: A low, creaky sound sometimes present in human speech
Factors That Can Affect Accuracy
Several factors can influence the accuracy of VoiceGuardAI's analysis:
Audio Quality
Poor audio quality can make it difficult to analyze the voice accurately. Factors that can affect quality include:
- Background noise
- Low recording volume
- Audio compression artifacts
- Distortion or clipping
Sample Length
Longer audio samples generally provide more accurate results:
- Less than 3 seconds: May not provide enough data for reliable analysis
- 3-10 seconds: Provides moderate reliability
- More than 10 seconds: Offers the most reliable analysis
AI Technology Advancements
As AI voice synthesis technology improves, detection becomes more challenging. VoiceGuardAI is continuously updated to keep pace with these advancements, but there's always an arms race between generation and detection technologies.
What to Do After Receiving Results
If Classified as "Likely Human"
Even if a voicemail is classified as likely human:
- Remain vigilant if the content seems suspicious
- Verify the caller's identity through other means if they're requesting sensitive information
- Be aware that very high-quality AI voices might occasionally be misclassified as human
If Classified as "Likely AI-Generated"
If a voicemail is classified as likely AI-generated:
- Do not return the call using the number provided in the voicemail
- Do not follow instructions in the voicemail, especially regarding financial transactions
- If the caller claims to be from a legitimate organization, contact that organization directly using their official contact information (not the number from the voicemail)
- Report the scam to relevant authorities
If Classified as "Uncertain"
If the result is uncertain:
- Treat the voicemail with caution
- Try scanning again with a better quality recording if possible
- Use additional verification methods before taking any action based on the voicemail
- Consider the context and content of the message - does it seem suspicious regardless of the voice?