Minimum Word Error Rate Training with Language Model Fusion for End-to-End Speech Recognition

Abstract

Integrating external language models (LMs) into end-to-end (E2E) models remains a challenging task for domain-adaptive speech recognition. Recently, internal language model estimation (ILME)-based LM fusion has shown significant word error rate (WER) reduction from Shallow Fusion by subtracting a weighted internal LM score from an interpolation of E2E model and external LM scores during beam search. However, on different test sets, the optimal LM interpolation weights vary over a wide range and have to be tuned extensively on well-matched validation sets. In this work, we perform LM fusion in the minimum WER (MWER) training of an E2E model to obviate the need for LM weights tuning during inference. Besides MWER training with Shallow Fusion (MWER-SF), we propose a novel MWER training with ILME (MWER-ILME) where the ILME-based fusion is conducted to generate N-best hypotheses and their posteriors. Additional gradient is induced when internal LM is engaged in MWER-ILME loss computation. During inference, LM weights pre-determined in MWER training enable robust LM integrations on test sets from different domains. Experimented with 30K-hour trained transformer transducers, MWER-ILME achieves on average 8.8% and 5.8% relative WER reductions from MWER and MWER-SF training, respectively, on 6 different test sets.

Sentence-Select: Large-Scale Language Model Data Selection for Rare-Word Speech Recognition

“…Text data for training the language model is often scraped from typed search logs to achieve the best domain match to smart assistant voice queries. These logs can be very large [2,4], making it prohibitively expensive to make even a single epoch through the data, limiting rare-word exposure. Furthermore, search queries can be heavy-headed, meaning that they contain disproportionately many high-frequency queries relative to low-frequency queries, also limiting rareword learning.…”

An Empirical Study of Language Model Integration for Transducer based Speech Recognition

“…All comparisons are made in decode-time manner, thus the works like [13,16,19,20] that modify the architecture or training objective of RNN-T are not included in this work. It should also be clarified that, when discussing "cross-domain", we assume the source domain and target domain are matched in acoustics, otherwise Eq.…”