Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 21·6 min read
CBT-Audio: Evaluating Audio Language Models for Patient-Side Distress Intensity Estimation in CBT Session Recordings
TL;DR
Cognitive behavioral therapy (CBT)—a talk-based mental health treatment—relies on therapists detecting emotional distress from vocal tone and speech patterns, but AI progress has been limited to text analysis. Researchers evaluated audio language models (neural networks trained on speech and language) on CBT-Audio, a dataset measuring patient distress intensity from session recordings.
✦ Why It Matters
Engineers building clinical AI tools now have a concrete benchmark and dataset to develop systems that detect patient distress from speech, not just transcripts.
Key Takeaways
- Audio language models outperform text-only approaches for detecting emotional distress in therapy.
- Paralinguistic features (tone, pace, hesitation) encode clinically relevant information text transcripts discard.
- CBT-Audio framework enables benchmarking of speech models on therapy distress detection tasks.
- Real-time distress detection from audio can alert therapists to escalating patient emotional states.
Full Summary
Cognitive behavioral therapy (CBT) is a psychological treatment where therapists listen to both content and delivery—vocal tone, hesitation, emotional intensity—to adapt their responses. Most AI systems supporting CBT work only with transcribed text, missing critical paralinguistic cues (emotional markers in speech itself) that therapists use clinically.
CBT-Audio introduces an evaluation framework that applies audio language models—neural networks trained on speech and language jointly—to estimate patient distress intensity directly from session recordings. The work addresses a gap where audio datasets for CBT were scarce, limiting development of speech-aware clinical AI.
Researchers evaluated multiple audio language model architectures on their dataset, measuring performance through metrics like correlation with clinician-rated distress scores. Results demonstrate that audio models can capture distress signals missed by text-only approaches, with implications for building AI systems that support therapists by flagging high-distress moments in real time.