Domain Adaptation Using Factorized Hidden Layer for Robust Automatic Speech Recognition

Sim, Khe Chai; Narayanan, Arun; Misra, Ananya; Tripathi, Anshuman; Pundak, Golan; Sainath, Tara N.; Haghani, Parisa; Li, Bo; Bacchiani, Michiel

doi:10.21437/interspeech.2018-2246

Cited by 32 publications

(19 citation statements)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Overfitting can be suppressed by using a lower learning rate and early stopping. This is consistent with the results reported in [12], where robust domain adaptation can be achieved by fine-tuning the entire model with millions of parameters on a small amount of adaptation data.…”

Section: Overfittingsupporting

confidence: 92%

Personalization of End-to-End Speech Recognition on Mobile Devices for Named Entities

Sim

Johnson

Motta

et al. 2019

2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)

Self Cite

View full text Add to dashboard Cite

We study the effectiveness of several techniques to personalize end-to-end speech models and improve the recognition of proper names relevant to the user. These techniques differ in the amounts of user effort required to provide supervision, and are evaluated on how they impact speech recognition performance. We propose using keyword-dependent precision and recall metrics to measure vocabulary acquisition performance. We evaluate the algorithms on a dataset that we designed to contain names of persons that are difficult to recognize. Therefore, the baseline recall rate for proper names in this dataset is very low: 2.4%. A data synthesis approach we developed brings it to 48.6%, with no need for speech input from the user. With speech input, if the user corrects only the names, the name recall rate improves to 64.4%. If the user corrects all the recognition errors, we achieve the best recall of 73.5%. To eliminate the need to upload user data and store personalized models on a server, we focus on performing the entire personalization workflow on a mobile device.

show abstract

Section: Overfittingsupporting

confidence: 92%

Personalization of End-to-End Speech Recognition on Mobile Devices for Named Entities

Sim

Johnson

Motta

et al. 2019

2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)

Self Cite

View full text Add to dashboard Cite

show abstract

“…There are a number of methods for adapting large ASR models to small amounts of data [12,13,14]. This paper's ap- Figure 1: Schematic diagrams of the RNN-T architecture (left) and the LAS architecture (right).…”

Section: Related Workmentioning

confidence: 99%

Personalizing ASR for Dysarthric and Accented Speech with Limited Data

Shor¹,

Emanuel²,

Lang³

et al. 2019

Interspeech 2019

View full text Add to dashboard Cite

Automatic speech recognition (ASR) systems have dramatically improved over the last few years. ASR systems are most often trained from typical speech, which means that underrepresented groups dont experience the same level of improvement. In this paper, we present and evaluate finetuning techniques to improve ASR for users with non-standard speech. We focus on two types of non-standard speech: speech from people with amyotrophic lateral sclerosis (ALS) and accented speech. We train personalized models that achieve 62% and 35% relative WER improvement on these two groups, bringing the absolute WER for ALS speakers, on a test set of message bank phrases, down to 10% for mild dysarthria and 20% for more serious dysarthria. We show that 71% of the improvement comes from only 5 minutes of training data. Finetuning a particular subset of layers (with many fewer parameters) often gives better results than finetuning the entire model. This is the first step towards building state of the art ASR models for dysarthric speech.

show abstract

“…The highest performance of speaker identification achieved 97.65% using GMFCC with 14 features employing YOHO database. Meanwhile, the accuracy using [30], reported an investigation into the use of Factorized Hidden Layer (FHL) to achieve compact model adaptation to unseen domains. The authors found that the SVD to initialize the low-rank bases of an FHL model leads to a faster convergence and improved performance.…”

Section: Accuracy =mentioning

confidence: 99%

Real and Complex Wavelet Transform Using Singular Value Decomposition for Malaysian Speaker and Accent Recognition

Abdullah

Vijean

Hariharan

et al. 2020

Lecture Notes in Mechanical Engineering

View full text Add to dashboard Cite

This paper presents a new approach for Malaysian speaker and accent recognition using wavelet feature extraction method, namely Wavelet Packet Transform (WPT), Discrete Wavelet Packet Transform (DWPT) and Dual Tree Complex Wavelet Packet Transform (DT-CWPT). Since Singular Value Decomposition (SVD) was based on factorization and summarization technique which reduces a rectangular matric, it is applied on those features to evaluate the performance for speaker and accent recognition. The features are derived from wavelets and SVD classified with three different classifiers namely k-Nearest Neighbors (k-NN), Support Vector Machine (SVM) and Extreme Learning Machine (ELM). In this work, English digits (0-9) and Malay words database uttered from 75 undergraduate students of Universiti Malaysia Perlis (UniMAP) which are Malays, Chinese and Indian. The Malay words had a combination of consonants and vowels in monosyllable and bi-syllable structure. The accuracy of file-based analysis achieved were above 81% while for frame-based analysis, 93.87% and above were obtained using three different classifiers (k-NN, SVM and ELM) for speaker and accent recognition. Through the experiments, it is observed that accent recognition achieved high recognition rate of 100% for both framed-based analysis and file-based analysis using SVM. The experimental results show the proposed features using SVD achieved high accuracy of 100% using SVM through English digits and Malay words in accent recognition. This indicated that feature extraction using wavelets (WPT, DWPT and DT-CWPT) with SVD can achieve a good performance for both English digits and Malay words.

show abstract

Domain Adaptation Using Factorized Hidden Layer for Robust Automatic Speech Recognition

Cited by 32 publications

References 32 publications

Personalization of End-to-End Speech Recognition on Mobile Devices for Named Entities

Personalization of End-to-End Speech Recognition on Mobile Devices for Named Entities

Personalizing ASR for Dysarthric and Accented Speech with Limited Data

Real and Complex Wavelet Transform Using Singular Value Decomposition for Malaysian Speaker and Accent Recognition

Contact Info

Product

Resources

About