Brief

A user on Reddit is seeking help with an Arabic Automatic Speech Recognition (ASR) model that is failing to converge during training. The model, based on a SpeechBrain Conformer-Transformer architecture, uses a combination of CTC and KL divergence loss functions. Despite significant drops in both losses early on, they quickly plateau, resulting in a high Word Error Rate (WER) on validation sets. The user has attempted various adjustments to learning rate, batch size, and vocabulary size without success, and is looking for potential causes or solutions from the community. AI

Researchers have developed an end-to-end neural decoder for intracortical speech decoding, aiming to eliminate the need for external language models. This Conformer-based system, trained on neural activity from an ALS patient, achieved a 23.80% character error rate without external linguistic support. The study indicates that signal degradation across sessions and word boundary segmentation are key challenges, but demonstrates the potential for a self-contained system to provide strong neural signals for speech processing. AI