The IBM 2015 English conversational telephone speech recognition system

By George Saon, Hong-Kwang Jeff Kuo, Steven J. Rennie, Michael A. Picheny. 2015

132 citations recorded by the source corpus.

We describe the latest improvements to the IBM English conversational telephone speech recognition system. Some of the techniques that were found beneficial are: maxout networks with annealed dropout rates; networks with a very large number of outputs trained on 2000 hours of data; joint modeling of partially unfolded recurrent neural networks and convolutional nets by combining the bottleneck and output layers and retraining the resulting model; and lastly, sophisticated language model rescoring with exponential and neural network LMs. These techniques result in an 8.0% word error rate on the Switchboard part of the Hub5-2000 evaluation test set which is 23% relative better than our previous best published result.

Subfields: Artificial Intelligence, 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 Signal Processing

Citations in the corpus

Achieving Human Parity in Conversational Speech Recognition (2016, background): “Surprisingly, this is the case for both acoustic modeling [28, 29, 30, 31, 32, 33] and language modeling [34, 35, 36, 37].”
Letter-Based Speech Recognition with Gated ConvNets (2017, background): “…end-to-end approaches leverage acoustic models (often ConvNet-based) topped with RNN layers as in (Hannun et al., 2014a; Miao et al., 2015; Saon et al., 2015; Amodei et al., 2016; Zhou et al., 2018; Zeyer et al., 2018) (e.g. a state of the art on WSJ (Chan and Lane, 2015a)), trained with…”
The microsoft 2016 conversational speech recognition system (2016, background): “They are similar in spirit to those studied more recently by [11, 10, 28].”
Comparing Human and Machine Errors in Conversational Speech Transcription (2017, methodology): “[7] demonstrated that deep networks as acoustic models could achieve significant improvements over GMM-HMM models on CTS data, and more recently researchers at IBM had achieved results on this task that represented a further significant advance [8, 9] over those from a decade ago.”
End-to-End attention based text-dependent speaker verification (2016, background): “This is inspired by recent works in [18, 19] which illustrated a deep convolution neural network (CNN) architecture outperformed LSTMs in many speech recognition tasks.”
DeepDefense: Identifying DDoS Attack via Deep Learning (2017, background): “RNN has achieved dramatic improvement in speech recognition [25], language translation [26], speech synthesis [27], and other sequential data [28] [29].”
Wav2Letter: an End-to-End ConvNet-based Speech Recognition System (2016, methodology): “There are now competitive end-to-end approaches of acoustic models toppled with RNNs layers as in [8, 13, 21, 1], trained with a sequence criterion [6].”
Improving End-to-End Speech Recognition with Policy Learning (2017, background): “Deep neural networks are the basis for some of the most accurate speech recognition systems in research and production [1, 2, 3].”
The Microsoft 2017 Conversational Speech Recognition System (2017, methodology): “The combination of deep networks and efficient training methods with older neural modeling concepts [2, 3, 4, 5, 6, 7, 8] have produced steady advances in both acoustic modeling [9, 10, 11, 12, 13, 14, 15] and language modeling [16, 17, 18, 19].”
The IBM 2016 Speaker Recognition System (2016, methodology): “The range of the LDA transformation is diagonalized by means of a global semi-tied covariance transform (see [20, 21] for more details).”
Purely Sequence-Trained Neural Networks for ASR Based on Lattice-Free MMI (2016, background): “Table 5: Comparison with systems in [22] and [23]”
Dense Prediction on Sequences with Time-Dilated Convolutions for Speech Recognition (2016, methodology): “SWB CH XE ST XE ST Classic 512 CNN [18] 12.6 10.4 IBM 2016 RNN+VGG+LSTM [19] 8.6 † 14.4 † MSR 2016 ResNet * [20] 8.9 MSR 2016 LACE * [20] 8.6 MSR 2016 BLSTM * [20] 8.7 VGG (pool, inefficient) [19] 10.2 9.4 16.3 16.0 VGG (no pool) [8] 10.8 9.7 17.1 16.7 VGG-10 + BN (no pool) [8] 10.8 9.5 17.0 16.3 VG”
The Technology Behind Personal Digital Assistants: An overview of the system architecture and key components (2017, background): “For voice queries, the first step is to recognize the spoken words [9], [10].”
Toward Human Parity in Conversational Speech Recognition (2017, result): “They are similar in spirit to those studied more recently by [30]–[33], [61].”