Short description

Electrolaryngeal (EL) speech presents unique acoustic challenges due to its artificial voicing and reduced prosodic variation. While state-of-the-art Automatic Speech Recognition (ASR) systems have demonstrated low Word Error Rates (WER) on standard datasets, their performance on pathological or EL speech in real-time conditions remains underexplored. This thesis investigates causal architectures suited for streaming applications. The study will evaluate state-of-the-art networks in streaming setups and evaluate their performance on pathological speech datasets.

Your Tasks

• Review current research in streaming ASR
• Benchmark ASR architectures (e.g., streaming Conformer, RNNT, Whisper with chunked decoding)
• Analyze latency, real-time factor, and robustness to EL-specific distortions
• Explore causal adaptation of models
• Document the methodologies, experimental setups, and results

Your Profile/Prerequisites

• Strong interest in ASR, speech technology, and pathological applications
• Experience with Python and ASR toolkits is beneficial (e.g., SpeechBrain, NVIDIA NeMo, OpenAI Whisper)
• Interests in causal inference and streaming architectures
• Familiarity with speech signal processing and speech communication

Contact:

• Martin Hagmüller (hagmueller@tugraz.at or 0316/873 4377)
• Benedikt Mayrhofer (benedikt.mayrhofer@tugraz.at)