Adapting Offline Speech Translation Models For Streaming With Future-aware Distillation And Inference

Abstract

A popular approach to streaming speech translation is to employ a single offline model with a wait-k policy to support different latency requirements, which is simpler than training multiple online models with different latency constraints. However, there is a mismatch problem in using a model trained with complete utterances for streaming inference with partial input. We demonstrate that speech representations extracted at the end of a streaming input are significantly different from those extracted from a complete utterance. To address this issue, we propose a new approach called Future-Aware Streaming Translation (FAST) that adapts an offline ST model for streaming input. FAST includes a Future-Aware Inference (FAI) strategy that incorporates future context through a trainable masked embedding, and a Future-Aware Distillation (FAD) framework that transfers future context from an approximation of full speech to streaming input. Our experiments on the MuST-C EnDe, EnEs, and EnFr bench

Stats

• citations2
• S2 citations—
• github stars0
• HF likes0
• heat score3.58
• arxiv keyfu2023adapting

Related papers

• Improving Streaming Automatic Speech Recognition With Non-streaming Model Distillation On Unsupervised Data (2020)0.00
• Streamatt: Direct Streaming Speech-to-text Translation With Attention-based Audio History Selection (2024)4.52
• Direct Simultaneous Speech-to-text Translation Assisted By Synchronized Streaming ASR (2021)6.77
• Stateful Conformer With Cache-based Inference For Streaming Automatic Speech Recognition (2023)8.60
• Teaching BERT To Wait: Balancing Accuracy And Latency For Streaming Disfluency Detection (2022)5.24
• Alignment Knowledge Distillation For Online Streaming Attention-based Speech Recognition (2021)7.16
• Learning When To Translate For Streaming Speech (2021)9.23
• Streaming Simultaneous Speech Translation With Augmented Memory Transformer (2020)6.77