Multi-style training for robust isolated-word speech recognition

By Richard P. Lippmann, E. A. Martin, D. Paul. 1987

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

1. Dong Yu
2. Jinyu Li
3. Li Deng
4. Bo Li
5. Tara N. Sainath
6. Shinji Watanabe
7. Yoshua Bengio
8. Yu Zhang

Citations in the corpus

• Exploring Speech Enhancement with Generative Adversarial Networks for Robust Speech Recognition (2017, background): “State-of-the-art ASR systems use MTR [21] to achieve robustness to noise at inference time.”
• A comparative study of traditional and newly proposed features for recognition of speech under stress (2000, background): “Others have also considered alternative training methods such as multi-style training [8].”
• Recent progresses in deep learning based acoustic models (2017, background): “For example, if the source domain is a clean environment and the target domain is a noisy environment, we can simulate noisy environment data and then do the multi-style training [159] with simulated data, e.g., in [160].”
• Data Augmentation for Deep Neural Network Acoustic Modeling (2015, background): “Digital Object Identi.er 10.1109/TASLP.2015.2438544 abundant variations presented in the training samples under the same labels and make classi.cations that are invariant to such variations.”
• Nonlinear feature based classification of speech under stress (2001, methodology): “similar method, called multi-style training [34], has been used to improve speech recognition under stress, but at the expense of requiring the user to produce speech across a simulated range of stress styles.”
• A study on data augmentation of reverberant speech for robust speech recognition (2017, background): “Multi-style training is a widely adopted strategy to train robust acoustic models [2].”
• Lower Frame Rate Neural Network Acoustic Models (2016, methodology): “For our experiments we use a mutli-style training procedure [17], where we first start with clean 3m utterances (about 2,000 hrs) which are anonymized and hand-transcribed voice search queries, and are representative of Google’s voice search traffic.”
• Large-Scale Domain Adaptation via Teacher-Student Learning (2017, methodology): “Knowledge distillation was also used to adapt an acoustic model to new dialects using a small adaptation corpus [17].”
• SpeakerStew: Scaling to Many Languages with a Triaged Multilingual Text-Dependent and Text-Independent Speaker Verification System (2021, methodology): “To improve robustness of the model, we apply data augmentation techniques involving noise sources combined with room simulation effects [25, 3, 26].”
• Automatic speech recognition and speech variability: A review (2007, background): “The same way, speech variations due to stress and emotions has been addressed by the multi-style training [169, 222], a nd simulated stress token generation [26, 27].”
• Speech recognition by machines and humans (1997, background): “Machine recognizers typically provide best performance with degraded speech only when they are trained using degraded speech materiŽ .als e.g. Lippmann and Martin, 1987 or when internal parameters are adapted to mimic this type of Ž .training e.g. Gopinath et al., 1995 .”
• The design and collection of COSINE, a multi-microphone in situ speech corpus recorded in noisy environments (2012, background): “The use of training audio that exhibits the Lombard effect has also been shown to improve the performance of speech recognition systems [7].”
• Domain Adaptation Using Factorized Hidden Layer for Robust Automatic Speech Recognition (2018, background): “This is referred to as multi-style training [14, 15, 16].”
• Single-channel mixed speech recognition using deep neural networks (2014, unknown): “One way to circumvent this issue is using a multi-style training strategy [16] in which training data is synthesized to be representative of the speech expected to be observed at test time.”