audio_detector.py

"""
Audio quality detector for transcription confidence.
"""

import logging
from typing import Dict, Any
from agentic_reliability_framework.runtime.agents.base import BaseAgent, AgentSpecialization
from ai_event import AIEvent

logger = logging.getLogger(__name__)

class AudioQualityDetector(BaseAgent):
    """
    Detects low‑confidence transcriptions based on average log probability.
    Expects `retrieval_scores[0]` to contain the average log probability.
    """

    def __init__(self):
        super().__init__(AgentSpecialization.DETECTIVE)
        self._threshold = -5.0  # average log prob threshold

    async def analyze(self, event: AIEvent) -> Dict[str, Any]:
        """
        Analyze audio transcription event.
        """
        avg_log_prob = event.retrieval_scores[0] if event.retrieval_scores else -10.0
        flags = []
        if avg_log_prob < self._threshold:
            flags.append('low_confidence')

        # Confidence derived from log prob (higher log prob → higher confidence)
        confidence = max(0, min(1.0, 1.0 + avg_log_prob / 10.0))

        return {
            'specialization': 'audio_quality',
            'confidence': confidence,
            'findings': {
                'flags': flags,
                'avg_log_prob': avg_log_prob
            },
            'recommendations': [
                'Use clearer audio',
                'Try a different model',
                'Check for background noise'
            ] if flags else []
        }