Explainability Subcard

Field	Response
Intended Task/Domain:	Speech Emotion Recognition, Audio Analysis, Human-Computer Interaction, and Audio2Face Integration
Model Type:	Speech emotion recognition classifier
Intended Users:	Audio2Face developers, Speech analysis researchers, Human-computer interaction developers, Affective computing researchers
Output:	Emotion probabilities (six classes: anger, disgust, fear, joy, neutral, and sadness)
Describe how the model works:	Audio input is processed through Wav2Vec2 architecture to classify emotions from speech, outputting probability scores for six emotional states
Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of:	People with speech disorders or non-native accents, Non-English speakers or those with strong regional accents, Elderly individuals with age-related speech changes
Technical Limitations & Mitigation:	Model requires clear audio input at 16kHz sampling rate, may struggle with overlapping speech or very noisy environments
Verified to have met prescribed NVIDIA quality standards:	Yes - Model achieves high accuracy on clean audio inputs, validated on internal crowdsourced dataset
Performance Metrics:	Accuracy (Top-1) - 80%+ on clean audio, Throughput & Latency, Emotion classification confidence scores
Potential Known Risks:	Model may misclassify emotions in edge cases, should not be used for standalone emotion analysis without Audio2Face integration
Licensing:	Use of this model is governed by the License Agreement for NVIDIA Audio2Emotion Model for Use with Audio2Face Project