Accurate and compact large vocabulary speech recognition on mobile devices

By X. Lei, Andrew W. Senior, Alexander Gruenstein, Jeffrey Scott Sorensen. 2013

128 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Efficiency & Compression, Speech & Audio, 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 Efficiency & Compression

In Speech & Audio

In Signal Processing

Citations in the corpus

Light Gated Recurrent Units for Speech Recognition (2018, background): “ASR, [70]–[75], which studies DNNs designed for portable devices with small computational capabilities.”
Generative Poisoning Attack Method Against Neural Networks (2017, background): “For example, a small-footprint large vocabulary speech recognizer for mobile devices [12], a small robust deep-learning model designed to provide high quality text-to-speech functionality on smart devices [3], and an audio sensing framework built from coupled DNNs [4].”
Clipper: A Low-Latency Online Prediction Serving System (2016, background): “Outside of content recommendation, there has been recent success in speech recognition [36, 55] and internet-scale resource allocation [23].”
Automatic gain control and multi-style training for robust small-footprint keyword spotting with deep neural networks (2015, methodology): “Our acoustic feature extraction module and the DNN runtime engine are both implemented using fixed-point arithmetic in order to minimize power consumption [18].”
Transferring knowledge from a RNN to a DNN (2015, methodology): “However, in narrow shallow networks, training with the same GMM alignments often hurts our ASR performance [11].”
MCDNN: An Approximation-Based Execution Framework for Deep Stream Processing Under Resource Constraints (2016, background): “For important models such as speech and object recognition, research teams have spent considerable effort producing manually optimized versions of individual DNNs that are efficient at run-time [32, 42, 45].”
Compressing deep neural networks using a rank-constrained topology (2015, methodology): “[13] F. Grézl and P. Fousek, “Optimizing bottle-neck features for LVCSR,” in Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2008, pp. 4729–4732.”
Improving DNN speaker independence with I-vector inputs (2014, background): “One of the problems with speaker adaptation is that it is hard to adapt a large number of parameters with only a small amount of data.”
Learning small-size DNN with output-distribution-based criteria (2014, methodology): “The accuracy gap is large between DNNs with large and small numbers of hidden nodes [11] if we train them with the standard training method in Section 2.”
Personalized speech recognition on mobile devices (2016, methodology): “In this paper we extend previous work that used quantized deep neural networks (DNNs) and on-the-ﬂy language model rescoring to achieve real-time performance on modern smartphones [1].”
On the compression of recurrent neural networks with an application to LVCSR acoustic modeling for embedded speech recognition (2016, methodology): “Although the traditional model for these applications has been to recognize speech remotely on large servers, there has been growing interest in developing ASR technologies that can recognize the input speech directly “on-device” [7].”
Wise teachers train better DNN acoustic models (2016, methodology): “[35] applied these optimizations along with frame skipping [36], which computes posteriors every nth frame and reuses them for n consecutive frames, to a small DNN for embedded recognition that ran in real-time with large accuracy improvements over a GMM acoustic model but still had a small footprin”
DEEProtect: Enabling Inference-based Access Control on Mobile Sensing Applications (2017, methodology): “The commonly used features for speech-to-text translation are Melfrequency Cepstral Coefficients (MFCC), spectrogram and Ngram [48], [61], [93], while the popular features for speaker identity recognition are MFCC, spectrogram and voice biometrics of the user (such as her pitch, accent, etc.”
Small-footprint keyword spotting using deep neural networks (2014, methodology): “For the HMM baseline system we use 5 future frames and 10 frames in the past, as this provided the best trade-off between accuracy, latency, and computation [15].”
Small-footprint high-performance deep neural network-based speech recognition using split-VQ (2015, background): “Index Terms— DNN, on device speech recognition, model compression, split-VQ”