Text-Utilization for Encoder-dominated Speech Recognition Models

arXiv (Cornell University)

Abstract

This paper investigates efficient methods for utilizing text-only data to improve speech recognition, focusing on encoder-dominated models that facilitate faster recognition. We provide a comprehensive comparison of techniques to integrate text-only data, including modality matching and dynamic downsampling to reach text-level representations within the encoder. Our experiments on the LibriSpeech corpus show that a larger encoder with a smaller decoder can equal or surpass the performance of architectures with larger decoders. We demonstrate that simple configurations, such as random duration models, are often more effective than complex alternatives, significantly simplifying the training pipeline. All code and recipes are made publicly available.

Authors 4

1. Albert Zeyer Aachen Machine Learning and Human Language Technology Group

   RWTH Aachen University

   Affiliation as printed

   AppTek , Germany

   Machine Learning and Human Language Technology Group , RWTH Aachen University , Germany

2. Tim Posielek Aachen Machine Learning and Human Language Technology Group

   RWTH Aachen University

   Affiliation as printed

   Machine Learning and Human Language Technology Group , RWTH Aachen University , Germany

3. Ralf Schlüter Aachen Machine Learning and Human Language Technology Group

   RWTH Aachen University

   Affiliation as printed

   AppTek , Germany

   Machine Learning and Human Language Technology Group , RWTH Aachen University , Germany

4. Hermann Ney Aachen Machine Learning and Human Language Technology Group

   RWTH Aachen University

   Affiliation as printed

   AppTek , Germany

   Machine Learning and Human Language Technology Group , RWTH Aachen University , Germany