A big data approach to acoustic model training corpus selection

By Olga Kapralova, J. Alex, Eugene Weinstein, Pedro J. Moreno Mengibar, Olivier Siohan. 2014

30 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. Dong Yu
3. Tara N. Sainath
4. Yu Zhang
5. Yoshua Bengio
6. Quoc V. Le
7. Bo Li
8. Yonghui Wu

In Speech & Audio

1. Jinyu Li
2. Tara N. Sainath
3. Yu Zhang
4. Ruoming Pang
5. Bo Li
6. Wei-Ning Hsu
7. Yonghui Wu
8. Abdel-rahman Mohamed

In Signal Processing

1. Jinyu Li
2. Dong Yu
3. Tara N. Sainath
4. Bo Li
5. Yu Zhang
6. Ruoming Pang
7. Awni Y. Hannun
8. Hasim Sak

Citations in the corpus

• Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin (2015, methodology): “ngine to align and filter a thousand hours of read speech [46]. In another approach, a heavy-weight offline speech recognizer is used to generate transcriptions for tens of thousands of hours of speech [33]. This is then passed through a filter and used to re-train the recognizer, resulting in significa”
• Self-Training for End-to-End Speech Recognition (2019, background): “Using pseudo-labels with a hybrid system can also give an improvement to WER on a large-scale data set [33] and in another, a student-teacher approach without an external LM gives an improvement when trained on unreleased unpaired audio [34].”
• Syllable-based acoustic modeling with CTC-SMBR-LSTM (2017, background): “9 million unsupervised high confidence utterances [18] from the same traffic.”
• Active and Semi-Supervised Learning in ASR: Benefits on the Acoustic and Language Models (2016, methodology): “As suggested in [11], utterance length and frequency filtering are additionally applied to flatten the data.”
• Semi-supervised Training for End-to-end Models via Weak Distillation (2019, methodology): “Confidence scores from an existing ASR system is commonly used to select unsupervised data for training with more data [11–15].”
• Neural Speech Recognizer: Acoustic-to-Word LSTM Model for Large Vocabulary Speech Recognition (2016, methodology): “While others have used read speech corpora [5, 22] or unsupervised methods [23] to gather thousands or even tens of thousands of hours of labeled training data, we apply an approach first described in [24] but now scaled up to build a training set of over 125,000 hours.”