Fast and Robust Unsupervised Contextual Biasing for Speech Recognition

Abstract

Automatic speech recognition (ASR) system is becoming a ubiquitous technology. Although its accuracy is closing the gap with that of human level under certain settings, one area that can further improve is to incorporate user-specific information or context to bias its prediction. A common framework is to dynamically construct a small language model from the provided contextual mini corpus and interpolate its score with the main language model during the decoding process.Here we propose an alternative approach that does not entail explicit contextual language model. Instead, we derive the bias score for every word in the system vocabulary from the training corpus. The method is unique in that 1) it does not require meta-data or class-label annotation for the context or the training corpus. 2) The bias score is proportional to the word's log-probability, thus not only would it bias the provided context, but also robust against irrelevant context (e.g. user misspecified or in case where it is hard to quantify a tight scope).3) The bias score for the entire vocabulary is pre-determined during the training stage, thereby eliminating computationally expensive language model construction during inference.We show significant improvement in recognition accuracy when the relevant context is available. Additionally, we also demonstrate that the proposed method exhibits high tolerance to false-triggering errors in the presence of irrelevant context.

Tree-constrained Pointer Generator for End-to-end Contextual Speech Recognition

“…Since there is no context vector in RNN-T, the query is derived from the encoder hidden state instead, as shown in Eqn. (13).…”

“…Contextual biasing, which integrates contextual knowledge into an automatic speech recognition (ASR) system, has become increasingly important to many applications [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18]. Contextual knowledge is often represented by a list (referred to as a biasing list) of words or phrases (referred to as biasing words) that are likely to appear in an utterance in a given context.…”

Tree-Constrained Pointer Generator for End-to-End Contextual Speech Recognition

“…Contextual speech recognition aims at addressing the long-tail word problem in end-to-end ASR systems by incorporating contextual knowledge, and has become increasingly important in many applications [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16]. Informative contextual knowledge can be extracted from text-based resources such as a users' contact book, playlist and the ontology in a dialogue system, or from other modalities such as presentation slides or objects in a scene.…”

“…Incorporating dynamic contextual knowledge into end-toend ASR systems is a challenging problem. Dedicated contextual biasing approaches have been developed, such as shallow fusion (SF) with a special weighted finite-state transducer (WFST), a language model (LM) adapted for contextual knowledge [1][2][3][14][15][16]; attention-based deep context approaches [4][5][6][7][8], and also deep biasing (DB) with a prefix-tree for improved efficiency when dealing with large biasing lists [9,10].…”

Spell my name: keyword boosted speech recognition

“…Contextual information has been researched by using fusion methods with trained language models (LM) [7,8,9], but also there are works dealing with contextual biasing which utilises a specific context such as named entity or personalised contacts. Majority of works require training an additional representation such as bias encoder [10,11,12], bias LM [13], class based LM [14,15,16], or additional data augmentation of a named entity with Text-to-speech [17]. Our work is similar to [18] that the bias information is encoded to trie structure.…”