Coupling Speech Encoders with Downstream Text Models
This addresses the challenge of improving speech translation quality when incremental training of the MT model is not possible, though it is incremental as the gain disappears with incremental training.
The paper tackles the problem of building cascade speech translation models that guarantee performance no worse than the 1-best baseline while preserving state-of-the-art ASR and MT performance, achieving this through an 'exporter' layer that matches ASR embeddings to MT token embeddings.
We present a modular approach to building cascade speech translation (AST) models that guarantees that the resulting model performs no worse than the 1-best cascade baseline while preserving state-of-the-art speech recognition (ASR) and text translation (MT) performance for a given task. Our novel contribution is the use of an ``exporter'' layer that is trained under L2-loss to ensure a strong match between ASR embeddings and the MT token embeddings for the 1-best sequence. The ``exporter'' output embeddings are fed directly to the MT model in lieu of 1-best token embeddings, thus guaranteeing that the resulting model performs no worse than the 1-best cascade baseline, while allowing back-propagation gradient to flow from the MT model into the ASR components. The matched-embeddings cascade architecture provide a significant improvement over its 1-best counterpart in scenarios where incremental training of the MT model is not an option and yet we seek to improve quality by leveraging (speech, transcription, translated transcription) data provided with the AST task. The gain disappears when the MT model is incrementally trained on the parallel text data available with the AST task. The approach holds promise for other scenarios that seek to couple ASR encoders and immutable text models, such at large language models (LLM).