Conversational privacy still seems like there is a lot of unsolved problems.
Problems
- How to timestamp audio and text corresponding to people?
- Are Privacy Preserving Phrases (P3) enough to determine privacy in context?
- If not, is emotion, intensity, volume, pitch, etc. an indicator of privacy (audio)?
- What is the best way to create a baseline model of general privacy vs non-privacy and fine-tune it to each user?
- How do we fine-tune the model to each user?
- What task do we present to the users in our studies?
Solutions
- I am sure there is some tool for this with deepspeech or google api
- ???
- ???
- Deep neural network trained on many private/non-private conversations. Another option could be using privacy scores or tags relating to each context or topic.
- We can have a configuration session, or do it during a tutorial. Alternatively, we can have preset privacy configurations depending on the users level of concern for privacy.
- Conversation memory bot? A robot playmate? A suite of Alexa apps?
Components
- Google natural language
- Privacy vs non-privacy conversation dataset
- Privacy vs non-privacy context triplets
- Privacy vs non-privacy DNN or scoring system
- Context tuples (topic, entity, sentiments, roles, activity, location) (emotion, intensity, volume, pitch)
- Speech transcription (Deepspeech)
- Speaker recognition
- Task software
