# Coupling Speech Encoders with Downstream Text Models

arXiv: https://arxiv.org/abs/2407.17605 · HTML: https://mlsanity.com/paper/2407.17605/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 11 of its 15 references that are in the map.
- First posted on arXiv 24 Jul 2024.

## Builds on (most cited first, 11 of 11)

- [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://mlsanity.com/paper/2104.09864/) (2021)
- [Neural Machine Translation by Jointly Learning to Align and Translate](https://mlsanity.com/paper/1409.0473/) (2014)
- [SentencePiece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing](https://mlsanity.com/paper/1808.06226/) (2018)
- [Conformer: Convolution-augmented Transformer for Speech Recognition](https://mlsanity.com/paper/2005.08100/) (2020)
- [LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech](https://mlsanity.com/paper/1904.02882/) (2019)
- [Exploring Architectures, Data and Units For Streaming End-to-End Speech Recognition with RNN-Transducer](https://mlsanity.com/paper/1801.00841/) (2018)
- [Searchable Hidden Intermediates for End-to-End Models of Decomposable Sequence Tasks](https://mlsanity.com/paper/2105.00573/) (2021)
- [Accelerating RNN-T Training and Inference Using CTC guidance](https://mlsanity.com/paper/2210.16481/) (2022)
- [Tight Integrated End-to-End Training for Cascaded Speech Translation](https://mlsanity.com/paper/2011.12167/) (2020)
- [LegoNN: Building Modular Encoder-Decoder Models](https://mlsanity.com/paper/2206.03318/) (2022)
- [Lego-Features: Exporting modular encoder features for streaming and deliberation ASR](https://mlsanity.com/paper/2304.00173/) (2023)

## Abstract (arXiv)

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).