docs/context/technical_spec.md

core components

• base model: Ethostral (fine-tuned Mistral).
• tracking and evaluation: weights and biases.
• platform: hugging face for model and adapter hosting.

architecture

1. process: audio is streamed to the fine-tuned mistral voxtral endpoint for simultaneous automatic speech recognition and emotion classification.
2. output format: transcription output uses interleaved text and emotional metadata tags.
3. frontend: Next.js application utilizing shadcn UI (Maia style) and Phosphor icons for the interactive dashboard.

integration points

• weights and biases weave: used for tracing the recognition pipeline.
• hugging face hub: serves as the repository for fine-tuned weights and dataset storage.
• shadcn ui: component library with maia theme.
• phosphor icons: primary iconography set.

performance metrics

• word error rate for transcription quality.
• f1 score for emotion detection accuracy.

benchmarking and evals

• IEMOCAP: Evaluation of categorical and dimensional (Valence/Arousal/Dominance) accuracy.
• RAVDESS: Benchmarking of prosodic feature mapping and speech rate accuracy.
• SUSAS: Evaluation of stress detection reliability under varied acoustic conditions.
• MDPE: Assessment of deception-related emotional leakage detection.
• Weights & Biases Weave: Used for tracking eval traces and scoring pipeline performance.