SLU for Voice Command in Smart Home: Comparison of Pipeline and End-to-End Approaches

Abstract

HAL is a multidisciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L'archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d'enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

SLURP: A Spoken Language Understanding Resource Package

“…Note that so far, the direct comparison of E2E-SLU with pipeline approaches are mainly limited to baselines developed on the same dataset, e.g. a multistage neural model in which the two stages that correspond to ASR and NLU are trained independently, but using the same training data (Desot et al, 2019;Haghani et al, 2018). We follow a different approach, which, as we argue, is closer to the 6 The official script for analysis and evaluation will be released with SLURP at https://github.com/ pswietojanski/slurp.…”

End-to-End Spoken Language Understanding: Performance analyses of a voice command task in a low resource setting

“…On top of that, our correlation tests in Section 5.3 showed that perfect ASR is not necessary to obtain good E2E SLU performance. It is, however, essential in the case of a pipeline approach as we have demonstrated in Desot et al (2019b) for intent prediction and in Desot et al (2019a) for concept prediction. This answers our first question and confirms the state-of-the-art: the E2E model reduces the cascade of error effect.…”

“…The E2E approach as outlined in Desot et al (2019a) was based on the ESPnet ASR toolkit (Watanabe et al, 2018). It integrates the Kaldi data preparation, extracts Mel filter-bank features, and combines Chainer and PyTorch deep learning tools (Tokui et al, 2015;Paszke et al, 2017).…”

“…For the E2E experiments, we used ESPnet default settings (Desot et al, 2019a) in order to train on speech data, with slots and intents symbolically injected in the transcriptions. Since the training was end-to-end, the training data was composed of the ASR training data (472.65 h) plus the NLU training data (81, 25) were the artificial dataset was synthesised (cf.…”

SLURP: A Spoken Language Understanding Resource Package

“…Note that so far, the direct comparison of E2E-SLU with pipeline approaches are mainly limited to baselines developed on the same dataset, e.g. a multistage neural model in which the two stages that correspond to ASR and NLU are trained independently, but using the same training data (Desot et al, 2019;Haghani et al, 2018). We follow a different approach, which, as we argue, is closer to the real-life application scenario: We use competitive ASR systems and state-of-the-art NLU systems.…”