Cognitive Speech Coding Milos Cernak, Senior Member, IEEE, Afsaneh Asaei, Senior Member, IEEE, Alexandre Hyafil

Cognitive Speech Coding Milos Cernak, Senior Member, IEEE, Afsaneh Asaei, Senior Member, IEEE, Alexandre Hyafil

1 Cognitive Speech Coding Milos Cernak, Senior Member, IEEE, Afsaneh Asaei, Senior Member, IEEE, Alexandre Hyafil Abstract—Speech coding is a field where compression ear and undergoes a highly complex transformation paradigms have not changed in the last 30 years. The before it is encoded efficiently by spikes at the auditory speech signals are most commonly encoded with com- nerve. This great efficiency in information representation pression methods that have roots in Linear Predictive has inspired speech engineers to incorporate aspects of theory dating back to the early 1940s. This paper tries to cognitive processing in when developing efficient speech bridge this influential theory with recent cognitive studies applicable in speech communication engineering. technologies. This tutorial article reviews the mechanisms of speech Speech coding is a field where research has slowed perception that lead to perceptual speech coding. Then considerably in recent years. This has occurred not it focuses on human speech communication and machine because it has achieved the ultimate in minimizing bit learning, and application of cognitive speech processing in rate for transparent speech quality, but because recent speech compression that presents a paradigm shift from improvements have been small and commercial applica- perceptual (auditory) speech processing towards cognitive tions (e.g., cell phones) have been mostly satisfactory for (auditory plus cortical) speech processing. The objective the general public, and the growth of available bandwidth of this tutorial is to provide an overview of the impact has reduced requirements to compress speech even fur- of cognitive speech processing on speech compression and discuss challenges faced in this interdisciplinary speech ther. However, better compression is always welcomed, processing field. In this context, it covers the traditional e.g. in large archival systems, etc. This article presents speech coding techniques as well as emerging approaches an overview of the basics of speech representation and facilitated by deep learning computational methods. The speech neuroscience, and outlines cognitively inspired tutorial points out key references on fundamental teachings speech coding that promises higher compression, adapt- of psycholinguistics and speech neuroscience and provides ability and robustness of the next generation of speech a valuable background to beginners and practitioners on coding technology. the promising directions of incorporating principles of Historically, the mechanisms of speech perception cognitive speech processing in speech compression. lead to perceptual speech coding [2], primarily suitable Index Terms—Speech coding, speech production and for digital audio. Substantial progress in this context perception, cognition, deep learning incorporates mechanisms to “optimize” coder perfor- mance for the human ear in the context of sub-band I. INTRODUCTION (transform) coders. On the other hand, the most common speech coding for medium to low bit rates is based Speech coding is an essential technology in informa- on models of human speech production, realized as tion transmission and communication systems. Human linear predictive vocoders, and analysis-by-synthesis lin- cognitive processing operates at about 50 bps (bits per ear predictive coders. Unified audio and speech coding second), which corresponds roughly to the speech pro- is usually realized with real-time switching according duction semantics as the rate of phonemic information to the input signal type. For example, the Enhanced in speech (e.g., most languages have approximately 32 Voice Services (EVS) coder standardized in 2015 by phonemes, encoded with 5 bits, and 1 s of speech has 3GPP offers new features and improvements for low- perhaps 10 phonemes), and the sensory system is known delay real-time communication systems, higher quality to encode non-redundant structures [1]. Efficient coding for clean/noisy speech, mixed content and music, in- maximizes the amount of information conveyed about cluding support for wideband, super-wideband and full- the sensory signal to the rest of the brain. The incoming band content [3]. However, the core speech compression acoustic signal is transmitted mechanically to the inner method is Algebraic Code Excited Linear Prediction Milos Cernak and Afsaneh Asaei are with Idiap Research Institute, (ACELP) proposed in 1987 by Adoul et al. [4]. The Centre du Parc, Rue Marconi 19, 1920 Martigny, Switzerland. Email: compression paradigm thus did not change significantly [email protected], [email protected]. in the last 30 years, and has its roots in Linear Predictive Alexandre Hyafil is with the Center for Brain and Cognition at Universitat Pompeu Babra, Barcelona, Spain. Email: alexan- theory dating back to the early 1940s [5], [6]. The dre.hyafi[email protected] significance of this tutorial is in bridging this influential 2 theory with cognitive studies highlight applicability to into units of different temporal granularity, such as speech communication engineering. phonemes and syllables, in parallel. As an example, the Research studies during the last decade incorporate incremental phonetic vocoder – cascaded speech recog- additional aspects of speech processing, namely func- nition and synthesis systems – extended with syllable- tional and temporal organization of human speech and based asynchronous information transmission mecha- language processing. Figure 1 shows the overall speech nisms was recently proposed [13]. The principles of perception process. Speech signal is treated indiscrim- asynchronous processing are fundamental in cortical inately by subcortical structures from other types of perception processing; asynchronicity exists in visual acoustic input. Such processes have now been fairly perception [14], in audiovisual perception [15], and in well characterized, particularly the cochlea where it is asynchronous evolution of various articulatory feature decomposed into different frequency channels forming, streams of speech recognition [16]. what has been coined, the auditory spectrogram [7]. The objective of this tutorial is to provide an overview Automatic speech processing is much less inspired from of the impact of cognitive speech processing on speech the subsequent stages in the auditory cortex, and in compression, and outline challenges faced in this inter- particularly omits how this continuous representation is disciplinary signal processing field. The article relates transformed into a discrete representation, i.e., a lexicon. to recent findings of speech and language neuroscience Sucha transformation is particularly difficult because the with traditional speech coding techniques in the context different units (phonemes, syllables, etc.) have varying of recent deep learning computational methods. The tuto- durations and operate at different time scales. rial points out key references on fundamental teachings The cognitive speech processing introduced in this of psycholinguistics and neurolinguistics and provides article, is specifically based on a dual-stream cortical a valuable background on the promising directions to circuit [8], and it presents a paradigm shift from per- incorporate principles of cognitive speech processing in ceptual (auditory) speech processing towards cognitive speech compression. (auditory/peripheral+cortical/central) sparse speech pro- Cognition in this article is assumed to be information cessing. Perception is extensively studied in general processing in the central nervous system, after the pe- auditory and speech processing [9]. Biologically, not ripheral auditory system. We avoid any broader definition only the cochlea but also the auditory cortex contribute to that relates to abstract concepts such as memory, mean- speech perception. Auditory sensations reach perception ing, mind and intelligence. Although cognition in speech only if received and processed by a cortical area. engineering might be understood as speech perception, Cognitive speech processing covers in particular the for clarity of our presentation, Section II assumes cog- temporal aspects of speech processing. As shown by nition as the underlying processes (algorithms) existing Fig. 1, the auditory cortex must deal with the different in both speech perception and production. Section III time scales pertaining to speech, and so one prominent briefly reviews perceptual audio coding and linear pre- hypothesis is that speech is first parsed into chunks diction speech coding in order to provide association corresponding to syllables and phonemes and then each with cognitive speech processing. The last Section IV chunk is categorized [10]. It has been shown that, introduces cognitive speech coding and compares its during speech processing, the brain generatesa cortical properties with linear predictive coding, concluding by oscillation in the θ-range (3-8 Hz) that may correspond outlining the challenges faced in this interdisciplinary to the syllable rate, and faster γ-range oscillations (25- speech compression field. 40 Hz) that correspond to the more transient acoustic properties. As a result, the fine (phonetic) structure of II. HUMAN COGNITIVE SPEECH PROCESSING the speech (the energy bursts underlying consonants) This section reviews the key results of encoding of have signal modulation even above 40 Hz. Although sound, and in particular speech sounds, by humans, and psychology and speech engineering already have a long its sparse and cortical representations. It is focused

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    15 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us