Large-vocabulary speech recognition under adverse acoustic environments

By Li Deng, A. Acero, M. Plumpe, Xuedong Huang. 2000

258 citations recorded by the source corpus.

We report our recent work on noise-robust large-vocabulary speech recognition. Three key innovations are developed and evaluated in this work: 1) a new model learning paradigm that comprises a noise-insertion process followed by noise reduction; 2) a noise adaptive training algorithm that integrates noise reduction into probabilistic multi-style system training; and 3) a new algorithm (SPLICE) for noise reduction that makes no assumptions about noise stationarity. Evaluation on a large-vocabulary speech recognition task demonstrates significant and consistent error rate reduction using these techniques. The resulting error rate is shown to be lower than that achieved by the matched-noisy condition for both stationary and nonstationary natural, as well as simulated, noises.

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

1. Li Deng
2. Dong Yu
3. Jinyu Li
4. Jianfeng Gao
5. Yoshua Bengio
6. Yifan Gong
7. Xiaodong He
8. Kyunghyun Cho

In Signal Processing

1. Li Deng
2. Dong Yu
3. Jinyu Li
4. Yifan Gong
5. S. Renals
6. M. Seltzer
7. Mark John Francis Gales
8. Jasha Droppo

Citations in the corpus

• Data Noising as Smoothing in Neural Network Language Models (2017, background): “Similarly, in speech recognition adding a background audio track or applying small shifts along the time dimension has been shown to yield significant gains, especially in noisy settings (Deng et al., 2000; Hannun et al., 2014).”
• An Investigation into the Effectiveness of Enhancement in ASR Training and Test for Chime-5 Dinner Party Transcription (2019, background): “Previous studies have shown the benefit of training an ASR on matching enhanced speech [6,7] or on jointly training the enhancement and the acoustic model [8].”
• Recent advances in deep learning for speech research at Microsoft (2013, background): “, [31][12]) to compensate for noise and channel mismatch.”
• Model-Based Approaches to Handling Uncertainty (2011, methodology): “As the KL divergence looks at complete distributions (rather than just the first-order moments in MMSE) changes in the uncertainty can be modelled with PCMLLR.”
• Speech Denoising and Dereverberation Using Probabilistic Models (2000, methodology): “For these experiments, we compared VSE to the SS algorithm described in [7].”
• Distributed speech processing in miPad's multimodal user interface (2002, methodology): “In this enhanced version of SPLICE, we not only minimize the static deviation from the clean to noisy cepstral vectors (as in the basic version of SPLICE), but also seek to minimize the dynamic deviation.”
• Speech-Centric Information Processing: An Optimization-Oriented Approach (2013, background): “KEYWORDS | Joint optimization; speech recognition; speechcentric information processing (SCIP); spoken language trans- lation (SLT); spoken language understanding (SLU); voice search”
• MiPad: a multimodal interaction prototype (2001, methodology): “Since this error increase is mainly due to the additive environment noise, the Dr Who CSR engine used our noise adaptive training [1] to improve the performance of the built -in microphone.”
• Machine Learning Paradigms for Speech Recognition: An Overview (2013, unknown): “[206], [208], [209], and vector Taylor series approximation”
• An investigation of deep neural networks for noise robust speech recognition (2013, methodology): “In the context of GMM-HMMs, this approach is referred to as feature-space noise adaptive training [20, 6] and this approach can be directly applied to DNN acoustic model.”
• Joint Uncertainty Decoding for Noise Robust Subspace Gaussian Mixture Models (2013, methodology): “Many feature domain approaches have been proposed including spectral subtraction, cepstral mean and variance normalization (CMN/CVN), cepstral maximum mean square error estimation [2], SPLICE [3], ALGONQUIN [4] and feature space vector Taylor series (VTS) compensation [5].”
• HMM adaptation using vector taylor series for noisy speech recognition (2000, background): “We can also pool training data from different environments [3].”
• Dynamic compensation of HMM variances using the feature enhancement uncertainty computed from a parametric model of speech distortion (2005, methodology): “This kind of “stereo” training data was also needed in our earlier work [6], which uses the SPLICE technique [4], [5] to compute the variance associated with speech feature enhancement.”
• High-performance robust speech recognition using stereo training data (2001, methodology): “In [3], we described a novel noise-reduction algorithm named SPLICE (Stereo-based Piecewise LInear Compensation for Environments) for the first time.”