thesis topics — speech & audio

• Automatic tool to monitor communication style
• Based on process communication model (PCM)
• In collaboration with Communicate 2 Connect
• Task: apply speech recognition technology for PCM:
  • non-verbal information (emotion, prosody)
  • key-words, key-phrases → frequency counts
  • spontaneous, noisy audio

• Approach
• interested partners: myforce, KBC, Sebeco, MatterSight, Communicate 2 Connect, VDAB, ...
• Full presentation

• Current machine learning techniques learn from labelled data
• Can only learn relations prepared by human annotators → costly
• Example: speech recognition
  • Works well for standard, prepared speech
  • Fails on new words, regiolects, dialects, hesitations, borken-off words, ...
• Aim: self-learning speech recognizer; weakly supervised

• Approaches: see how humans do this and mimic it
• Hawkins' memory-prediction framework (On Intelligence: How a New Understanding of the Brain will Lead to the Creation of Truly Intelligent Machines)
• Deep Neural Networks (DNNs) with attention mechanisms

• Speech therapy for patients with severe speech impairment: laryngectomy, dysarthria
• Tool (website) to automate voice training (in collaboration with speech therapists)
• Web-site: recording (audio) + manual annotations (annotations) + automatic feedback

• Feedback to
  • patient (showing the progress he/she made):         
  • speech therapist (help diagnosing problems, adjust the exercises)
• Problems:
  • high variability (languages, type of speech disorder, age of the speaker)
  • lack of annotated data
  • handling spontaneous speech (speak freely)
• Solutions:
  • use of data from different languages?
  • use normal speech and convert to alaryngeal or dysarthric domain?
  • search similar audio in existing data?

• In collaboration with  
• More info: see lifewatch.be
• Goal: custom-made real time data processing tool to recognize bat species (or groups)
• Approach: exemplar-based techniques

• In collaboration with  
• More info: see website Flanders Marine Institute
• Goals: new classification algorithms to detect sounds from other marine mammals and even boat sonars; open algorithms to allow other more versatile hydrophones in the network

• Detect when somebody sings ↔ instruments only
  Non-trivial problem
  → deep learning