Do End-to-End Speech Recognition Models Care About Context?

Borgholt, Lasse; Havtorn, Jakob D.; Agić, Željko; Søgaard, Anders; Maaløe, Lars; Igel, Christian

doi:10.21437/interspeech.2020-1750

Cited by 9 publications

(6 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The performance of CTC significantly improved by using Conformer (B1, C1) and Mask-CTC greatly benefited from it (C2). The errors were further reduced by applying DLP (C3), achieving 9.1% in eval92 which was the best among the NAR models and better than that of the state-of-the-art model without LM [46,47]. By comparing results between NAR and AR models, Mask-CTC achieved highly competitive performance to AR models for both Transformer (A1, B3) and Conformer (A3, C3), demonstrating the effectiveness of the proposed methods for improving the original Mask-CTC.…”

Section: Resultsmentioning

confidence: 99%

Improved Mask-CTC for Non-Autoregressive End-to-End ASR

Higuchi

Inaguma

Watanabe

et al. 2021

ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

For real-world deployment of automatic speech recognition (ASR), the system is desired to be capable of fast inference while relieving the requirement of computational resources. The recently proposed end-to-end ASR system based on mask-predict with connectionist temporal classification (CTC), Mask-CTC, fulfills this demand by generating tokens in a non-autoregressive fashion. While Mask-CTC achieves remarkably fast inference speed, its recognition performance falls behind that of conventional autoregressive (AR) systems. To boost the performance of Mask-CTC, we first propose to enhance the encoder network architecture by employing a recently proposed architecture called Conformer. Next, we propose new training and decoding methods by introducing auxiliary objective to predict the length of a partial target sequence, which allows the model to delete or insert tokens during inference. Experimental results on different ASR tasks show that the proposed approaches improve Mask-CTC significantly, outperforming a standard CTC model (15.5% → 9.1% WER on WSJ). Moreover, Mask-CTC now achieves competitive results to AR models with no degradation of inference speed (< 0.1 RTF using CPU). We also show a potential application of Mask-CTC to end-to-end speech translation.

show abstract

Section: Resultsmentioning

confidence: 99%

Improved Mask-CTC for Non-Autoregressive End-to-End ASR

Higuchi

Inaguma

Watanabe

et al. 2021

ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

show abstract

“…ASR is a deep neural network using a model based on Connectionist Temporal Classification. 28 To train the ASR model in the Swedish language, a total of 45 h of uniform random selected Swedish emergency calls from 2015 concerning all types of emergencies were manually transcribed into written text. All text files were then used as labels for the ASR model to be trained in understanding the Swedish language.…”

Section: Automatic Speech Recognitionmentioning

confidence: 99%

Machine learning can support dispatchers to better and faster recognize out-of-hospital cardiac arrest during emergency calls: A retrospective study

et al. 2021

View full text Add to dashboard Cite

Aim: Fast recognition of out-of-hospital cardiac arrest (OHCA) by dispatchers might increase survival. The aim of this observational study of emergency calls was to (1) examine whether a machine learning framework (ML) can increase the proportion of calls recognizing OHCA within the first minute compared with dispatchers, (2) present the performance of ML with different false positive rate (FPR) settings, (3) examine call characteristics influencing OHCA recognition.Methods: ML can be configured with different FPR settings, i.e., more or less inclined to suspect an OHCA depending on the predefined setting. ML OHCA recognition within the first minute is evaluated with a 1.5 FPR as the primary endpoint, and other FPR settings as secondary endpoints. ML was exposed to a random sample of emergency calls from 2018. Voice logs were manually audited to evaluate dispatchers time to recognition.Results: Of 851 OHCA calls, the ML recognized 36% (n = 305) within 1 min compared with 25% (n = 213) by dispatchers. The recognition rate at any time during the call was 86% for ML and 84% for dispatchers, with a median time to recognition of 72 versus 94 s. OHCA recognized by both ML and dispatcher showed a 28 s mean difference in favour of ML (P < 0.001). ML with higher FPR settings reduced recognition times. Conclusion:ML recognized a higher proportion of OHCA within the first minute compared with dispatchers and has the potential to be a supportive tool during emergency calls. The optimal FPR settings need to be evaluated in a prospective study.

show abstract

“…For experiments with fine-tuning, we use language-specific BERT models 11 for German (Chan et al, 2020), Spanish (Canete et al, 2020), Dutch (de Vries et al, 2019, Finnish (Virtanen et al, 2019), Danish, 12 Croatain (Ulčar and Robnik-Šikonja, 2020), while we use mBERT (Devlin et al, 2019) for Afrikaans.…”

Section: Low Resource Settingmentioning

confidence: 99%

Context-aware Adversarial Training for Name Regularity Bias in Named Entity Recognition

Ghaddar

Langlais

Ahmad

et al. 2021

Transactions of the Association for Computational Linguistics

View full text Add to dashboard Cite

In this work, we examine the ability of NER models to use contextual information when predicting the type of an ambiguous entity. We introduce NRB, a new testbed carefully designed to diagnose Name Regularity Bias of NER models. Our results indicate that all state-of-the-art models we tested show such a bias; BERT fine-tuned models significantly outperforming feature-based (LSTM-CRF) ones on NRB, despite having comparable (sometimes lower) performance on standard benchmarks. To mitigate this bias, we propose a novel model-agnostic training method that adds learnable adversarial noise to some entity mentions, thus enforcing models to focus more strongly on the contextual signal, leading to significant gains on NRB. Combining it with two other training strategies, data augmentation and parameter freezing, leads to further gains.

show abstract

Do End-to-End Speech Recognition Models Care About Context?

Cited by 9 publications

References 30 publications

Improved Mask-CTC for Non-Autoregressive End-to-End ASR

Improved Mask-CTC for Non-Autoregressive End-to-End ASR

Machine learning can support dispatchers to better and faster recognize out-of-hospital cardiac arrest during emergency calls: A retrospective study

Context-aware Adversarial Training for Name Regularity Bias in Named Entity Recognition

Contact Info

Product

Resources

About