Voice mode transcription should use AI context to improve accuracy

Problem

The voice/listening mode transcription quality is poor. It appears to transcribe words individually without considering the context of the conversation, the topic being discussed, or sentence structure.

Examples

• Transcribes "from listening" as "from music" — showing it doesn't understand the surrounding context
• Frequently used words in the conversation are not recognized better over time
• The transcription just guesses words in isolation rather than understanding what the subject is

Suggestion

The transcription should use AI/LLM context to improve accuracy:

1. Topic awareness — Recognize what the conversation is about and use that context to disambiguate words (e.g., if discussing "voice mode", prefer "listening" over "music")
2. Sentence structure — Consider the full sentence and grammar, not just individual words in isolation
3. Frequent word learning — Track which words and terms are used often in the session and weight those higher in transcription
4. Subject-aware decoding — Use the conversation history and topic to inform the speech-to-text model's output

The current behavior feels like raw speech-to-text without any post-processing or contextual refinement. Adding an AI layer that understands the structure and subject of the conversation would significantly improve transcription quality.