Ralf Schlüter

Gradient-based alignment can outperform traditional methods in challenging scenarios, offering a universal solution for precise speech-to-text mapping across diverse ASR models.

The traditional linear correlation between language model perplexity and ASR word error rate breaks down in modern end-to-end systems, revealing the critical role of internal language modeling.

Large output values from positional encodings can severely compromise the performance of memristor-based computations, but targeted adjustments can halve this degradation.

Diffusion language models can substantially boost speech recognition accuracy, rivaling traditional language models while offering unique advantages like bidirectional attention.

Forget simply bolting on an LLM: this work reveals the surprisingly intricate dance between acoustic models and LLMs needed to unlock state-of-the-art speech recognition.

Unsupervised speech recognition is possible under specific theoretical conditions, paving the way for training models without paired data.