Generation of Large-Scale Simulated Utterances in Virtual Rooms to Train Deep-Neural Networks for Far-Field Speech Recognition in Google Home

By Chanwoo Kim, Ananya Misra, Kean K. Chin, Thad Hughes, Arun Narayanan, Tara N. Sainath, Michiel Bacchiani. 2017

247 citations recorded by the source corpus.

Subfields: Artificial Intelligence, 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

1. Jinyu Li
2. Yu Zhang
3. Yonghui Wu
4. Tara N. Sainath
5. Dong Yu
6. Shinji Watanabe
7. Wei-Ning Hsu
8. Ye Jia

In Speech & Audio

1. Yu Zhang
2. Jinyu Li
3. Yonghui Wu
4. Tara N. Sainath
5. Wei-Ning Hsu
6. Ye Jia
7. Ron J. Weiss
8. Chung-Cheng Chiu

In Signal Processing

1. Jinyu Li
2. Yu Zhang
3. Tara N. Sainath
4. Yonghui Wu
5. Dong Yu
6. Shinji Watanabe
7. Bo Li
8. Chung-Cheng Chiu

Citations in the corpus

• Robust wav2vec 2.0: Analyzing Domain Shift in Self-Supervised Pre-Training (2021, neutral): “This includes data augmentation [13, 14], self-training on target domain [15], domain adversarial training [16], and joint training [8].”
• SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition (2019, neutral): “The use of an acoustic room simulator has been explored in [14].”
• Toward Domain-Invariant Speech Recognition via Large Scale Training (2018, methodology): “The reverberation time is set to be between 0 and 900 msec, with a target to mic distance between 1 and 10 meters (see [3] for details).”
• Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesis (2018, methodology): “To achieve this, we pass the single-speaker US English proprietary dataset from (Wang et al., 2017a) into a room simulator (Kim et al., 2017), which adds varying types of background noise and room reverberations.”
• Exploring Speech Enhancement with Generative Adversarial Networks for Robust Speech Recognition (2017, methodology): “As a source of reverberation for MTR, we use a room simulator as described in [27].”
• Hierarchical Generative Modeling for Controllable Speech Synthesis (2018, methodology): “As a first experiment, we artificially generated training sets using a room simulator (Kim et al., 2017) to add background noise and reverberation to clean speech from the multi-speaker English corpus used in the previous section.”
• Streaming End-to-end Speech Recognition for Mobile Devices (2018, methodology): “This data set is created by artificially corrupting clean utterances using a room simulator, adding varying degrees of noise and reverberation such that the overall SNR is between 0dB and 30dB, with an average SNR of 12dB [44].”
• Specaugment on Large Scale Datasets (2019, methodology): “The work [3] studies the effect of using acoustic room simulators.”
• Snips Voice Platform: an embedded Spoken Language Understanding system for private-by-design voice interfaces (2018, methodology): “Thousands of virtual rooms of different sizes are thus generated with random microphone and speaker locations, and the rerecording of the original data in those conditions is simulated using a method close to that presented in [22].”
• Training Keyword Spotting Models on Non-IID Data with Federated Learning (2020, methodology): “Two common speech data augmentation methods—MTR [37] and SpecAugment [38]—were tuned for non-IID training.”
• Domain Adaptation Using Factorized Hidden Layer for Robust Automatic Speech Recognition (2018, background): “This is referred to as multi-style training [14, 15, 16].”
• Speech Processing for Digital Home Assistants: Combining signal processing with deep-learning techniques (2019, methodology): “Additional data augmentation techniques [32] are used in the development of an AM for a new application, which perturb existing recordings along different perceptually relevant parameters, such as speaking rate, vocal tract length, pitch, SNR, noise types, and so forth.”
• Multi-Dialect Speech Recognition with a Single Sequence-to-Sequence Model (2017, methodology): “using a room simulator, adding varying degrees of noise and reverberation such that the overall SNR is between 0 and 20dB [31].”
• An Attention-Based Joint Acoustic and Text on-Device End-To-End Model (2020, methodology): “This data set is created by artificially corrupting clean utterances using a room simulator, adding varying degrees of noise and reverberation such that the overall SNR is between 0dB and 30dB, with an average SNR of 12dB [26].”
• Convolutional Recurrent Neural Network and Data Augmentation for Audio Tagging with Noisy Labels and Minimal Supervision (2019, background): “It has been shown to improve classifier performance on many tasks, including speech recognition [6] and audio classification [7, 8, 9].”
• No Need for a Lexicon? Evaluating the Value of the Pronunciation Lexica in End-to-End Models (2017, methodology): “This data set is created by artificially corrupting clean utterances using a room simulator, adding varying degrees of noise and reverberation such that the overall SNR is between 0dB and 30dB, with an average SNR of 12dB [17].”
• Recognizing Long-Form Speech Using Streaming End-to-End Models (2019, methodology): “Techniques such as multi-condition training have been shown to improve performance in noisy and farfield conditions [29, 30].”