Skip to main content

Streaming Punctuation for Long-Form Dictation with Transformers



Streaming Punctuation for Long-Form Dictation with Transformers

Authors

Piyush Behre, Sharman Tan, Padma Varadharajan and Shuangyu Chang, Microsoft Corporation, USA

Abstract

While speech recognition Word Error Rate (WER) has reached human parity for English, longform dictation scenarios still suffer from segmentation and punctuation problems resulting from irregular pausing patterns or slow speakers. Transformer sequence tagging models are effective at capturing long bi-directional context, which is crucial for automatic punctuation. Automatic Speech Recognition (ASR) production systems, however, are constrained by real-time requirements, making it hard to incorporate the right context when making punctuation decisions. In this paper, we propose a streaming approach for punctuation or re-punctuation of ASR output using dynamic decoding windows and measure its impact on punctuation and segmentation accuracy across scenarios. The new system tackles over-segmentation issues, improving segmentation F0.5-score by 13.9%. Streaming punctuation achieves an average BLEU-score improvement of 0.66 for the downstream task of Machine Translation (MT).

Keywords

Automatic punctuation, automatic speech recognition, re-punctuation, speech segmentation.

Comments

Post a Comment

Popular posts from this blog

Comparison of Support Vector Machines and Deep Learning for Plant Classification in Smart Agriculture Applications

Comparison of Support Vector Machines and Deep Learning for Plant Classification in Smart Agriculture Applications Authors Esmael Hamuda 1, Ashkan Parsi 2, Martin Glavin 2 and Edward Jones 2, 1 Elmergib University, Libya, 2 University of Galway, Ireland Abstract In this paper, we investigate the use of deep learning approaches for plant classification (cauliflower and weeds) in smart agriculture applications. To perform this, five approaches were considered, two based on well-known deep learning architectures (AlexNet and GoogleNet), and three based on Support Vector Machine (SVM) classifiers with different feature sets (Bag of Words in L*a*b colour space, Bag of Words in HSV colour space, Bag of Words of Speeded-up Robust Features (SURF)). Two types of datasets were used in this study: one without Data Augmentation and the second one with Data Augmentation. Each algorithm's performance was tested with one data set similar to the training data, and a second data set acquired under ...
Post a Comment
Read more

Submit your Research Article - International Journal of Chaos, Control, Modelling and Simulation (IJCCMS)

Submit your Research Article!! International Journal of Chaos, Control, Modelling and Simulation (IJCCMS) ISSN :  2319 - 5398 [Online] ; 2319 - 8990 [Print] Webpage URL:  https://airccse.org/journal/ijccms/index.html Submission URL:  http://coneco2009.com/submissions/imagination/home.html Here's where you can reach us :  ijccmsjournal@yahoo.com or ijccms@aircconline.com
Post a Comment
Read more

MESSAGE EMBEDDED CIPHER USING 2-D CHAOTIC MAP

  Paper Title: MESSAGE EMBEDDED CIPHER USING 2-D CHAOTIC MAP Authors: Mina Mishra1and Dr. V.H. Mankar2, 1,2 Nagpur University, Nagpur, Maharashtra, India ABSTRACT This paper constructs two encryption methods using 2-D chaotic maps, Duffings and Arnold’s cat maps respectively. Both of the methods are designed using message embedded scheme and are analyzed for their validity, for plaintext sensitivity, key sensitivity, known plaintext and brute-force attacks. Due to the less key space generally many chaotic cryptosystem developed are found to be weak against Brute force attack which is an essential issue to be solved. For this issue, concept of identifiability proved to be a necessary condition to be fulfilled by the designed chaotic cipher to resist brute force attack, which is a basic attack. As 2-D chaotic maps provide more key space than 1-D maps thus they are considered to be more suitable. This work is accompanied with analysis results obtained from these developed cipher. Mor...
Post a Comment
Read more