scholarly journals Performance and Efficiency Evaluation of ASR Inference on the Edge

2021 ◽  
Vol 13 (22) ◽  
pp. 12392
Author(s):  
Santosh Gondi ◽  
Vineel Pratap
Keyword(s):  
Energy Consumption ◽  
Speech Recognition ◽  
Error Rate ◽  
Network Connectivity ◽  
Raspberry Pi ◽  
Word Error Rate ◽  
Trade Off ◽  
Network Connection ◽  
Long Run ◽  
User Data

Automatic speech recognition, a process of converting speech signals to text, has improved a great deal in the past decade thanks to the deep learning based systems. With the latest transformer based models, the recognition accuracy measured as word-error-rate (WER), is even below the human annotator error (4%). However, most of these advanced models run on big servers with large amounts of memory, CPU/GPU resources and have huge carbon footprint. This server based architecture of ASR is not viable in the long run given the inherent lack of privacy for user data, reliability and latency issues of the network connection. On the other hand, on-device ASR (meaning, speech to text conversion on the edge device itself) solutions will fix deep-rooted privacy issues while at same time being more reliable and performant by avoiding network connectivity to the back-end server. On-device ASR can also lead to a more sustainable solution by considering the energy vs. accuracy trade-off and choosing right model for specific use cases/applications of the product. Hence, in this paper we evaluate energy-accuracy trade-off of ASR with a typical transformer based speech recognition model on an edge device. We have run evaluations on Raspberry Pi with an off-the-shelf USB meter for measuring energy consumption. We conclude that, in the case of CPU based ASR inference, the energy consumption grows exponentially as the word error rate improves linearly. Additionally, based on our experiment we deduce that, with PyTorch mobile optimization and quantization, the typical transformer based ASR on edge performs reasonably well in terms of accuracy and latency and comes close to the accuracy of server based inference.

scholarly journals Dynamic Acoustic Unit Augmentation with BPE-Dropout for Low-Resource End-to-End Speech Recognition

Sensors ◽  
2021 ◽  
Vol 21 (9) ◽  
pp. 3063
Author(s):  
Aleksandr Laptev ◽  
Andrei Andrusenko ◽  
Ivan Podluzhny ◽  
Anton Mitrofanov ◽  
Ivan Medennikov ◽  
...  
Keyword(s):  
Speech Recognition ◽  
Error Rate ◽  
Rapid Development ◽  
Computational Cost ◽  
Vocabulary Size ◽  
Word Error Rate ◽  
Low Resource ◽  
Steady Improvement ◽  
End To End ◽  
Asr System

With the rapid development of speech assistants, adapting server-intended automatic speech recognition (ASR) solutions to a direct device has become crucial. For on-device speech recognition tasks, researchers and industry prefer end-to-end ASR systems as they can be made resource-efficient while maintaining a higher quality compared to hybrid systems. However, building end-to-end models requires a significant amount of speech data. Personalization, which is mainly handling out-of-vocabulary (OOV) words, is another challenging task associated with speech assistants. In this work, we consider building an effective end-to-end ASR system in low-resource setups with a high OOV rate, embodied in Babel Turkish and Babel Georgian tasks. We propose a method of dynamic acoustic unit augmentation based on the Byte Pair Encoding with dropout (BPE-dropout) technique. The method non-deterministically tokenizes utterances to extend the token’s contexts and to regularize their distribution for the model’s recognition of unseen words. It also reduces the need for optimal subword vocabulary size search. The technique provides a steady improvement in regular and personalized (OOV-oriented) speech recognition tasks (at least 6% relative word error rate (WER) and 25% relative F-score) at no additional computational cost. Owing to the BPE-dropout use, our monolingual Turkish Conformer has achieved a competitive result with 22.2% character error rate (CER) and 38.9% WER, which is close to the best published multilingual system.

scholarly journals Is a Wizard-of-Oz Required for Robot-Led Conversation Practice in a Second Language?

2022 ◽  
Author(s):  
Olov Engwall ◽  
José Lopes ◽  
Ronald Cumbal
Keyword(s):  
Second Language ◽  
Speech Recognition ◽  
Statistical Method ◽  
Error Rate ◽  
State Of The Art ◽  
Autonomous Robot ◽  
Language Learner ◽  
Word Error Rate ◽  
Wizard Of Oz ◽  
Custom Made

AbstractThe large majority of previous work on human-robot conversations in a second language has been performed with a human wizard-of-Oz. The reasons are that automatic speech recognition of non-native conversational speech is considered to be unreliable and that the dialogue management task of selecting robot utterances that are adequate at a given turn is complex in social conversations. This study therefore investigates if robot-led conversation practice in a second language with pairs of adult learners could potentially be managed by an autonomous robot. We first investigate how correct and understandable transcriptions of second language learner utterances are when made by a state-of-the-art speech recogniser. We find both a relatively high word error rate (41%) and that a substantial share (42%) of the utterances are judged to be incomprehensible or only partially understandable by a human reader. We then evaluate how adequate the robot utterance selection is, when performed manually based on the speech recognition transcriptions or autonomously using (a) predefined sequences of robot utterances, (b) a general state-of-the-art language model that selects utterances based on learner input or the preceding robot utterance, or (c) a custom-made statistical method that is trained on observations of the wizard’s choices in previous conversations. It is shown that adequate or at least acceptable robot utterances are selected by the human wizard in most cases (96%), even though the ASR transcriptions have a high word error rate. Further, the custom-made statistical method performs as well as manual selection of robot utterances based on ASR transcriptions. It was also found that the interaction strategy that the robot employed, which differed regarding how much the robot maintained the initiative in the conversation and if the focus of the conversation was on the robot or the learners, had marginal effects on the word error rate and understandability of the transcriptions but larger effects on the adequateness of the utterance selection. Autonomous robot-led conversations may hence work better with some robot interaction strategies.

scholarly journals Indigenuous Vocabulary Reformulation for Continuousyorùbá Speech Recognition In M-Commerce Using Acoustic Nudging-Based Gaussian Mixture Model

2021 ◽  
Author(s):  
Kehinde Lydia Ajayi ◽  
Victor Azeta ◽  
Isaac Odun-Ayo ◽  
Ambrose Azeta ◽  
Ajayi Peter Taiwo ◽  
...  
Keyword(s):  
Speech Recognition ◽  
Gaussian Mixture Model ◽  
Mixture Model ◽  
Error Rate ◽  
System Performance ◽  
Recognition Rate ◽  
Gaussian Mixture ◽  
Computer Applications ◽  
Word Error Rate ◽  
The Mean

Abstract One of the current research areas is speech recognition by aiding in the recognition of speech signals through computer applications. In this research paper, Acoustic Nudging, (AN) Model is used in re-formulating the persistence automatic speech recognition (ASR) errors that involves user’s acoustic irrational behavior which alters speech recognition accuracy. GMM helped in addressing low-resourced attribute of Yorùbá language to achieve better accuracy and system performance. From the simulated results given, it is observed that proposed Acoustic Nudging-based Gaussian Mixture Model (ANGM) improves accuracy and system performance which is evaluated based on Word Recognition Rate (WRR) and Word Error Rate (WER)given by validation accuracy, testing accuracy, and training accuracy. The evaluation results for the mean WRR accuracy achieved for the ANGM model is 95.277% and the mean Word Error Rate (WER) is 4.723%when compared to existing models. This approach thereby reduce error rate by 1.1%, 0.5%, 0.8%, 0.3%, and 1.4% when compared with other models. Therefore this work was able to discover a foundation for advancing current understanding of under-resourced languages and at the same time, development of accurate and precise model for speech recognition.

scholarly journals Exploiting Memory Resilience for Emerging Technologies: An Energy-Aware Resilience Exemplar for STT-RAM Memories

2020 ◽  
pp. 505-526
Author(s):  
Amir Mahdi Hosseini Monazzah ◽  
Amir M. Rahmani ◽  
Antonio Miele ◽  
Nikil Dutt
Keyword(s):  
Energy Consumption ◽  
Error Rate ◽  
Spin Transfer Torque ◽  
Error Correction Codes ◽  
Energy Aware ◽  
Trade Off ◽  
Protection Scheme ◽  
And Performance ◽  
On Chip ◽  
Rate Threshold

AbstractDue to the consistent pressing quest of larger on-chip memories and caches of multicore and manycore architectures, Spin Transfer Torque Magnetic RAM (STT-MRAM or STT-RAM) has been proposed as a promising technology to replace classical SRAMs in near-future devices. Main advantages of STT-RAMs are a considerably higher transistor density and a negligible leakage power compared with SRAM technology. However, the drawback of this technology is the high probability of errors occurring especially in write operations. Such errors are asymmetric and transition-dependent, where 0 → 1 is the most critical one, and is high subjected to the amount and current (voltage) supplied to the memory during the write operation. As a consequence, STT-RAMs present an intrinsic trade-off between energy consumption vs. reliability that needs to be properly tuned w.r.t. the currently running application and its reliability requirement. This chapter proposes FlexRel, an energy-aware reliability improvement architectural scheme for STT-RAM cache memories. FlexRel considers a memory architecture provided with Error Correction Codes (ECCs) and a custom current regulator for the various cache ways and conducts a trade-off between reliability and energy consumption. FlexRel cache controller dynamically profiles the number of 0 → 1 transitions of each individual bit write operation in a cache block and based on that selects the most-suitable cache way and current level to guarantee the necessary error rate threshold (in terms of occurred write errors) while minimizing the energy consumption. We experimentally evaluated the efficiency of FlexRel against the most efficient uniform protection scheme from reliability, energy, area, and performance perspectives. Experimental simulations performed by using gem5 has demonstrated that while FlexRel satisfies the given error rate threshold, it delivers up to 13.2% energy saving. From the area footprint perspective, FlexRel delivers up to 7.9% cache ways’ area saving. Furthermore, the performance overhead of the FlexRel algorithm which changes the traffic patterns of the cache ways during the executions is 1.7%, on average.

scholarly journals Konversi Suara Ucapan Bahasa Indonesia Ke Sistem Bahasa Isyarat Indonesia (Sibi)

2020 ◽  
Vol 2 (2) ◽  
pp. 7-13
Author(s):  
Andi Nasri
Keyword(s):  
Speech Recognition ◽  
Automatic Speech Recognition ◽  
Error Rate ◽  
Word Error Rate ◽  
Bahasa Indonesia

Dengan semakin berkembangnya teknologi speech recognition, berbagai software yang bertujuan untuk memudahkan orang tunarungu dalam berkomunikasi dengan yang lainnya telah dikembangkan. Sistem tersebut menterjemahkan suara ucapan menjadi bahasa isyarat atau sebaliknya bahasa isyarat diterjemahkan ke suara ucapan. Sistem tersebut sudah dikembangkan dalam berbagai bahasa seperti bahasa Inggris, Arab, Spanyol, Meksiko, Indonesia dan lain-lain. Khusus untuk bahasa Indonesia mulai juga sudah yang mencoba melakukan penelitian untuk membuat system seperti tersebut. Namun system yang dibuat masih terbatas pada Automatic Speech Recognition (ASR) yang digunakan dimana mempunyai kosa-kata yang terbatas. Dalam penelitian ini bertujuan untuk mengembangkan sistem penterjemah suara ucapan bahasa Indonesia ke Sistem Bahasa Isyarat Indonesia (SIBI) dengan data korpus yang lebih besar dan meggunkanan continue speech recognition  untuk meningkatkan akurasi system.Dari hasil pengujian system menunjukan diperoleh hasil akurasi sebesar rata-rata 90,50 % dan Word Error Rate (WER)  9,50%. Hasil akurasi lebih tinggi dibandingkan penelitian kedua  48,75%  dan penelitan pertama 66,67%. Disamping itu system juga dapat mengenali kata yang diucapkan secara kontinyu atau pengucapan kalimat. Kemudian hasil pengujian kinerja system mencapai         0,83 detik untuk Speech to Text  dan 8,25 detik untuk speech to sign.

scholarly journals Minimum Word Error Rate Training with Language Model Fusion for End-to-End Speech Recognition

2021 ◽  
Author(s):  
Zhong Meng ◽  
Yu Wu ◽  
Naoyuki Kanda ◽  
Liang Lu ◽  
Xie Chen ◽  
...  
Keyword(s):  
Speech Recognition ◽  
Error Rate ◽  
Language Model ◽  
Word Error Rate ◽  
Model Fusion ◽  
End To End

scholarly journals Building Acoustic and Language Model for Continuous Speech Recognition in Bahasa Indonesia

2020 ◽  
Vol 6 (2) ◽  
Author(s):  
Vincent Elbert Budiman ◽  
Andreas Widjaja
Keyword(s):  
Speech Recognition ◽  
Error Rate ◽  
Language Model ◽  
Beam Width ◽  
Acoustic Model ◽  
Continuous Speech ◽  
Continuous Speech Recognition ◽  
Word Error Rate ◽  
Testing Data ◽  
Bahasa Indonesia

Here a development of an Acoustic and Language Model is presented. Low Word Error Rate is an early good sign of a good Language and Acoustic Model. Although there are still parameters other than Words Error Rate, our work focused on building Bahasa Indonesia with approximately 2000 common words and achieved the minimum threshold of 25% Word Error Rate. There were several experiments consist of different cases, training data, and testing data with Word Error Rate and Testing Ratio as the main comparison. The language and acoustic model were built using Sphinx4 from Carnegie Mellon University using Hidden Markov Model for the acoustic model and ARPA Model for the language model. The models configurations, which are Beam Width and Force Alignment, directly correlates with Word Error Rate. The configurations were set to 1e-80 for Beam Width and 1e-60 for Force Alignment to prevent underfitting or overfitting of the acoustic model. The goals of this research are to build continuous speech recognition in Bahasa Indonesia which has low Word Error Rate and to determine the optimum numbers of training and testing data which minimize the Word Error Rate.  

scholarly journals Đánh giá các hệ thống nhận dạng giọng nói tiếng việt (vais, viettel, zalo, fpt và google) trong bản tin

2021 ◽  
pp. 28-35
Author(s):  
Nguyen Thi My Thanh ◽  
Phan Xuan Dung ◽  
Nguyen Ngoc Hay ◽  
Le Ngoc Bich ◽  
Dao Xuan Quy
Keyword(s):  
Speech Recognition ◽  
Automatic Speech Recognition ◽  
Error Rate ◽  
Viet Nam ◽  
Word Error Rate

Bài báo này giới thiệu kết quả đánh giá các hệ thống nhận dạng giọng nói tiếng Việt (VASP-Vietnamese Automatic Speech Recognition) trong bản tin từ các công ty hàng đầu của Việt Nam như Vais (Vietnam AI System), Viettel, Zalo, Fpt và công ty hàng đầu thế giới Google. Để đánh giá các hệ thống nhận dạng giọng nói, chúng tôi sử dụng hệ số Word Error Rate (WER) với đầu vào là văn bản thu được từ các hệ thống Vais VASP, Viettel VASP, Zalo VASP, Fpt VASP và Google VASP. Ở đây, chúng tôi sử dụng tập tin âm thanh là các bản tin và API từ các hệ thống Vais VASP, Viettel VASP, Zalo VASP, Fpt VASP và Google VASP để đưa ra văn bản được nhận dạng tương ứng. Kết quả so sánh WER từ Vais, Viettel, Zalo, Fpt và Google cho thấy hệ thống nhận dạng tiếng nói tiếng Việt trong các bản tin từ Viettel, Zalo, Fpt và Google đều có kết quả tốt, trong đó Vais cho kết quả vượt trội hơn.

Sign in / Sign up

Export Citation Format

Share Document