Weighted finite-state transducers in speech recognition

By Mehryar Mohri, Fernando C Pereira, Michael D. Riley. 2002

1,142 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Natural Language Processing, Speech & Audio

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

1. Dong Yu
2. Jinyu Li
3. Yoshua Bengio
4. Li Deng
5. Quoc V. Le
6. Tara N. Sainath
7. Shinji Watanabe
8. James R. Glass

In Natural Language Processing

1. Yoshua Bengio
2. Dong Yu
3. Graham Neubig
4. Kyunghyun Cho
5. Quoc V. Le
6. Jianfeng Gao
7. Yonatan Belinkov
8. Li Deng

In Speech & Audio

1. Jinyu Li
2. Shinji Watanabe
3. Tara N. Sainath
4. Dong Yu
5. Yu Zhang
6. Wei-Ning Hsu
7. James R. Glass
8. Yoshua Bengio

Citations in the corpus

• Unsupervised Speech Recognition (2021, methodology): “In practice, decoding is often carried out by searching the minimal distortion path on a weighted finite state transducer (WFST; Mohri 1997; Mohri et al. 2002), obtained by composing WFSTs representing HMM emission, HMM transition, pronunciation model, and language model, respectively.”
• Iterative Pseudo-Labeling for Speech Recognition (2020, neutral): “External Language Model (LM) In modern ASR systems, in addition to the acoustic model pθ , a decoding procedure (typically either WFST-based [22] or beam-search-based (BS) [23, 24]) integrates an external language model (LM).”
• End-to-end attention-based large vocabulary speech recognition (2015, methodology): “We use the Weighted Finite State Transducer (WFST) framework [18, 19] to build a character-level language model from a word-level one.”
• On Multiplicative Integration with Recurrent Neural Networks (2016, methodology): “Decoding is done with the CTC [21] based weighted finite-state transducers (WFSTs) [22] as proposed by [18].”
• EESEN: End-to-end speech recognition using deep RNN models and WFST-based decoding (2015, methodology): “The lexicon and grammar WFSTs are firstly composed.”
• Streaming End-to-end Speech Recognition for Mobile Devices (2018, methodology): “To construct the contextual LM, we assume that a set of wordlevel biasing phrases are known ahead of time, and compile them into a weighted finite state transducer (WFST) [38].”
• An Analysis of Incorporating an External Language Model into a Sequence-to-Sequence Model (2017, methodology): “Following [2], an alternative is to train the LM at the word level, and then, using the Weighted Finite State Transducer framework [17, 18], compose it with a “speller” which breaks each word into its component units (graphemes or wordpieces).”
• Snips Voice Platform: an embedded Spoken Language Understanding system for private-by-design voice interfaces (2018, methodology): “The order in which the compositions are done is important, as the composition (H ∗ C ∗ L) ∗G is known to be intractable when G is not deterministic, as is the case when G is a wFST representing an n-gram model [34].”
• A Bottom-Up Modular Search Approach to Large Vocabulary Continuous Speech Recognition (2013, background): “On the other hand, the minimization and determinazation are not performed to the final recognition network generated by (3), as commonly done for ASR system implemented with WFSMs [24], [59].”
• Quaternion Recurrent Neural Networks (2018, methodology): “Decoding is based on Kaldi [26] and weighted finite state transducers (WFST) [19] that integrate acoustic, lexicon and language model probabilities into a single HMM-based search graph.”
• Composition-based on-the-fly rescoring for salient n-gram biasing (2015, methodology): “It is at this point where we apply an on-the-fly composition [6, 7] with the above WFST representing the set of n-grams to bias.”
• A Pruned Rnnlm Lattice-Rescoring Algorithm for Automatic Speech Recognition (2018, background): “The reader can refer to [12] for a more precise explanation, but the basic algorithm is as follows.”
• Low-Latency Real-Time Meeting Recognition and Understanding Using Distant Microphones and Omni-Directional Camera (2012, methodology): “The decoder is based on WFSTs for speech recognition [33].”
• Fast, Simpler and More Accurate Hybrid ASR Systems Using Wordpieces (2020, methodology): “These posterior distributions are then fed into a weighted finite-state transducer (WFST) [33] based decoder with a decoding graph also composed with a lexicon and language model (LM) to find the best hypothesis.”
• A fast maximum likelihood nonlinear feature transformation method for GMM-HMM speaker adaptation (2014, methodology): “We use Kaldi [20] to train a weighted finite state transducer (WFST)-based recognition system [21].”
• Large-Scale Visual Speech Recognition (2018, methodology): “As described earlier, our model produces a sequence of phoneme distribution followed by standard decoding using finite state transducers (FSTs) to arrive at word sequences [42, 43].”