Bayesian Language Model Interpolation for Mobile Speech Input

By Cyril Allauzen, Michael D. Riley. 2011

54 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Large Language Models, Speech & Audio, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Large Language Models

In Speech & Audio

In Signal Processing

Citations in the corpus

Toward Domain-Invariant Speech Recognition via Large Scale Training (2018, methodology): “The LM is a Bayesian interpolated 5-gram model trained from a variety of data sources like anonymized and aggregated search queries and dictated texts [24].”
State-of-the-Art Speech Recognition with Sequence-to-Sequence Models (2017, methodology): “Since domains have different predictive value for our LVCSR task, domain-specific LMs are first trained, then combined together using Bayesian-interpolation [26].”
Federated Learning for Mobile Keyboard Prediction (2018, methodology): “The primary (static) language model for the English language in Gboard is a Katz smoothed Bayesian interpolated [4] 5-gram LM containing 1.”
Exploring architectures, data and units for streaming end-to-end speech recognition with RNN-transducer (2017, methodology): “The language model is pruned to 100-million n-grams with a target vocabulary of 4 million and the various sources of text data are re-weighted using interpolation [25] for the optimal word error rate performance.”
An Analysis of Incorporating an External Language Model into a Sequence-to-Sequence Model (2017, methodology): “In line 5 of Table 3, LAS-WP + RNN-WP, we show that the shallow fusion of LAS with RNN-WP provides an even greater benefit than PRODLM2.”
Domain Adaptive Inference for Neural Machine Translation (2019, methodology): “This corresponds to the approach in Allauzen and Riley (2011), which considers only language model combination for speech recognition.”
Improved recognition of contact names in voice commands (2015, methodology): “We used the Bayesian interpolation method of [6] to combine the voice command models with LMs trained on our other training sources.”
Personalized speech recognition on mobile devices (2016, methodology): “We compare performance obtained from the baseline system to a Bayesian interpolated LM [23], where voice commands and dictation are each represented as a unique task and the corresponding task priors are determined by sweeping parameters on a held-out development set to minimize word error rates…”