Snips Voice Platform: an embedded Spoken Language Understanding system for private-by-design voice interfaces

By Alice Coucke, Alaa Saade, Adrien Ball, Théodore Bluche, A. Caulier, David Leroy, Clément Doumouro, Thibault Gisselbrecht, F. Caltagirone, Thibaut Lavril and 2 more. 2018

921 citations recorded by the source corpus.

This paper presents the machine learning architecture of the Snips Voice Platform, a software solution to perform Spoken Language Understanding on microprocessors typical of IoT devices. The embedded inference is fast and accurate while enforcing privacy by design, as no personal user data is ever collected. Focusing on Automatic Speech Recognition and Natural Language Understanding, we detail our approach to training high-performance Machine Learning models that are small enough to run in real-time on small devices. Additionally, we describe a data generation procedure that provides sufficient, high-quality training data without compromising user privacy.

Subfields: Artificial Intelligence, Natural Language Processing, Speech & Audio, Privacy & Security, Signal Processing

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

In Natural Language Processing

In Speech & Audio

In Privacy & Security

In Signal Processing

Citations in the corpus

Federated Learning for Keyword Spotting (2018, background): “Once the wake word is detected, voice input is activated and processed by a spoken language understanding engine, powering the perception abilities of the voice assistant [1].”
Lifelong Intent Detection via Multi-Strategy Rebalancing (2021, methodology): “Based on four widely used datasets, ATIS [10], SNIPS [5], HWU64 [16], CLINC150 [14], we constructed four benchmarks.”
An Evaluation Dataset for Intent Classification and Out-of-Scope Prediction (2019, background): “(Li and Roth, 2002) 6 5,952 TREC-50, (Li and Roth, 2002) 50 5,952 Web Apps, (Braun et al., 2017) 8 89 Ask Ubuntu, (Braun et al., 2017) 5 162 Chatbot Corpus, (Braun et al., 2017) 2 206 Snips, (Coucke et al., 2018) 7 14,484 Liu et al. (2019) 54 25,716 Table 4: Classiﬁcation dataset properties.”
SUPERB: Speech processing Universal PERformance Benchmark (2021, methodology): “Audio SNIPS [18] is adopted, which synthesized multi-speaker utterances for SNIPS [30].”
Neural Data Augmentation via Example Extrapolation (2021, methodology): “We apply Ex2 to a range of language understanding tasks and significantly improve over state-of-the-art methods on multiple few-shot learning benchmarks, including for relation extraction (FewRel) and intent classification + slot filling (SNIPS).”
Data Augmentation using Pre-trained Transformer Models (2020, background): “Label Names SST-2 Positive, Negative TREC Description, Entity, Abbreviation, Human, Location, Numeric SNIPS PlayMusic, GetWeather, RateBook, SearchScreeningEvent, SearchCreativeWork, AddTo-Playlist, BookRestaurant For our experiments, we used the labels provided in Table 1.”
Unknown Intent Detection Using Gaussian Mixture Model with an Application to Zero-shot Intent Classification (2020, methodology): “We evaluate our method SEG for unknown intent detection on 3 real task-oriented dialogue datasets: SNIPS (Coucke et al., 2018), ATIS (Hemphill et al., 1990) and SMP-2018 (Zhang et al., 2017).”
Slot-Gated Modeling for Joint Slot Filling and Intent Prediction (2018, methodology): “This dataset is collected from the Snips personal voice assistant, where the number of samples for each intent is approximately the same (Coucke et al., 2018).”
Semi-Supervised Spoken Language Understanding via Self-Supervised Speech and Language Model Pretraining (2020, methodology): “INTRODUCTION Spoken Language Understanding (SLU)(1) is at the front-end of many modern intelligent home devices, virtual assistants, and socialbots [1, 2]: given a spoken command, an SLU engine should extract relevant semantics(2) from spoken commands for the appropriate downstream tasks.”
BiToD: A Bilingual Multi-Domain Dataset For Task-Oriented Dialogue Modeling (2021, background): “Different from aforementioned datasets focus on dialogue level modeling, there are many SLU datasets [40, 41, 42, 43] have been proposed for utterance level semantic parsing.”
Meta-learning for Few-shot Natural Language Processing: A Survey (2020, background): “SNIPS (Coucke et al., 2018) 4 is an intent classi-ﬁcation dataset with only seven intent types.”
Few-Shot Named Entity Recognition: A Comprehensive Study (2020, background): “…2017 (Derczynski et al., 2017) on social domain, MIT Moive (Liu et al., 2013b) and MIT Restaurant 2 (Liu et al., 2013a) on review domain, SNIPS 3 (Coucke et al., 2018), ATIS 4 (Hakkani-Tür et al., 2016) and Multi-woz 5 (Budzianowski et al., 2018) on dialogue domain, and I2B2 6 (Stubbs and…”
Composed Variational Natural Language Generation for Few-shot Intents (2020, methodology): “For SNIPS-NLU, the performance is calculated with the average and the standard deviation over 5 runs.”
DeepStruct: Pretraining of Language Models for Structure Prediction (2022, methodology): “Datasets We use two datasets, the ATIS dataset (Hemphill et al., 1990), which contains ﬂight and airline-related conversation and queries, and the SNIPS dataset (Coucke et al., 2018), which consists of daily queries from the interaction between the users and dialogue agents.”
A Closer Look At Feature Space Data Augmentation For Few-Shot Intent Classification (2019, methodology): “We compare these six FDA techniques on two open datasets for Intent Classification (IC) : SNIPS (Coucke et al., 2018) and Facebook Dialog corpus (Gupta et al., 2018).”
BERT for Joint Intent Classification and Slot Filling (2019, methodology): “More recently, a pre-training technique, Bidirectional Encoder Rep-resentations from Transformers (BERT) (Devlin et al., 2018), was proposed and has created state-of-the-art models for a wide variety of NLP tasks, including question answering (SQuAD v1.”