Speech Recognition with Weighted Finite-State Transducers

By Mehryar Mohri, Fernando C Pereira, Michael D. Riley. 2008

342 citations recorded by the source corpus.

View Paper: https://www.semanticscholar.org/paper/ea20537a9a73aeab2b97fdededbae57bd4f663cf

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

Citations in the corpus

Differentiable Weighted Finite-State Transducers (2020, background): “We can alleviate these issues with pruning and back-off (Katz, 1987) which can be encoded in a WFST (Mohri et al., 2008).”
Search by voice in Mandarin Chinese (2010, methodology): “We use standard decision-tree state-based clustering followed by semi-tied co-variance (STC) modeling [5] and an FST-based search [6].”
Semi-Supervised Training of Acoustic Models Using Lattice-Free MMI (2018, methodology): “Using the same method as in [2], we process these lattices into phone graphs that are compiled into utterance-specific FSTs by composing the graphs with a context-dependency transducer (C) [19, 20] and transducer representing the HMM topology (H).”
Personalized speech recognition on mobile devices (2016, methodology): “An FST-based decoder graph for the CTC model is created by the usual construction and composition of lexicon and LM transducers [24].”
Improving Transformer-Based End-to-End Speech Recognition with Connectionist Temporal Classification and Language Model Integration (2019, methodology): “, alphabet, Japanese character, etc) IDs Y [1 : u] = Y [1], .”
Lattice decoding and rescoring with long-Span neural network language models (2014, background): “By contrast, for transducer-based decoders, the sum over all reachable scores is taken into account instead, and it is even found that a look ahead based on the single best path can decrease performance ([20]).”
Mobile Keyboard Input Decoding with Finite-State Transducers (2017, background): “General background on FST decoding for speech can be found in [6].”
A Pruned Rnnlm Lattice-Rescoring Algorithm for Automatic Speech Recognition (2018, methodology): “We describe the algorithm in the weighted finite state transducer (FST) framework commonly used in most of the major speech recognition systems [11].”
High-performance brain-to-text communication via handwriting (2021, methodology): “Word frequencies are then encoded into a weighted finite state transducer (Mohri et al., 2008) that can be used to infer the most likely sequence of characters given the word frequency counts combined with the character probabilities output by the neural decoder.”
Snips Voice Platform: an embedded Spoken Language Understanding system for private-by-design voice interfaces (2018, methodology): “trained on the resulting set of patterns, which is then converted to a wFST called Gp [35].”
Purely Sequence-Trained Neural Networks for ASR Based on Lattice-Free MMI (2016, background): “The construction of the denominator graph from the phone-level denominator LM can be summarized as follows (see [11] for notation).”
Improved recognition of contact names in voice commands (2015, methodology): “We use the weight pushing algorithm [4] to redistribute the weights along the path while preserving full-path weights in such a way that at every non-initial state, the ⊕log-sum of the weights of all outgoing transitions is 0 (1 in the probability domain).”
No Need for a Lexicon? Evaluating the Value of the Pronunciation Lexica in End-to-End Models (2017, background): “Typically, the process of finding the best set of recognized words in the network is represented as a composition of finite state transducers (FSTs) [14], shown by Equation 1.”
Hybrid Autoregressive Transducer (HAT) (2020, background): “The phonetic model structure, pronunciations and LM are commonly represented and combined as weighted finite-state transducers (WFSTs) [3, 4].”