The Challenges of Testing Voice-Controlled Audio Systems

Smart devices that are voice-controlled such as smart speakers, hearables, and vehicle infotainment systems are notoriously complex to test. They have numerous connections from wired to wireless and contain much signal processing, both on the record and the playback side. This means that their characteristics change according to ‘real world’ conditions of the environment that they are used in, such as background noise, playback levels, and room acoustics. Furthermore, their multifunctional nature means that there are many aspects of the device that may need to be tested, ranging from voice recognition to music playback, operation as a hands-free telephone, and in the case of hearables, hearing assistance. Due to their complex non-linear use cases, these devices often need to be tested at different levels and different environmental conditions. This paper focuses on tools and techniques to accurately measure the audio performance of such devices under the many various real-world conditions in which they are used.

 

语音控制的智能设备(例如智能扬声器、听觉设备和车辆信息娱乐系统)非常难以测试。它们具有从有线到无线的多样连接方式,并且在接收端和重放端使用了诸多信号处理技术。这意味着它们的特性会随着使用环境的“现实世界”条件(例如背景噪声、播放级别和室内声学条件)的不同而变化。 此外,它们的多功能特性意味着可能需要测试该设备的许多方面,包括语音识别、音乐播放、作为免提电话或听觉设备或助听器使用时的性能。由于其复杂的非线性使用情况,这些设备通常需要在不同级别和不同环境条件下进行测试。本文重点介绍在各种实际条件下准确测量此类设备的音频性能的工具和技术。

Author: Steve Temme, Listen, Inc.
Presented at ISEAT 2019, Shenzhen, China.

Full Paper – English Version
Full Paper – Chinese Version

Testing Audio Performance of Hearables

Smart headphones or “hearables” are designed not only to playback music but to enhance communications in the
presence of background noise and in some cases, even compensate for hearing loss. They may also provide voice
recognition, medical monitoring, fitness tracking, real-time translation and even augmented reality (AR). They
contain complex signal processing and their characteristics change according to their smartphone application and
‘real world’ conditions of their actual environment, including background noises and playback levels. This paper
focuses on how to measure their audio performance under the many various real-world conditions they are used
in.

Authors: Steve Temme, Listen, Inc.
Presented at AES Headphone Conference 2019, San Francisco, CA.

Full Paper
Poster Presentation

A New THD+N Algorithm for Measuring Today’s High Resolution Audio Systems

In this paper, a mathematical definition of Total Harmonic Distortion + Noise suitable for testing high-resolution digital audio systems is presented. This formal definition of the “distortion analyzer” mentioned in AES17 defines THD+N as the RMS error of fitting a sinusoid to a noisy and distorted sequence of measurements. We present the key theoretical result that under realistic conditions a modern THD+N analyzer is well-described by a Normal probability distribution with a simple relationship between relative error and analysis dwell time. These findings are illustrated by comparing the output of a commercial distortion analyzer to our proposed method using Monte Carlo simulations of noisy signal channels. We will demonstrate that the bias of a well-designed distortion analyzer is negligible.

Authors: Alfred B. Roney The Mathworks, Inc. (formerly Listen, Inc.) Steve Temme, Listen, Inc.
Presented at AES 2018, New York, NY.

Full Paper

Evaluation of audio test methods and measurements for end-of-line loudspeaker quality control

In order to minimize costly warranty repairs, loudspeaker OEMS impose tight specifications and a “total quality” requirement on their part suppliers. At the same time, they also require low prices. This makes it important for driver manufacturers and contract manufacturers to work with their OEM customers to define reasonable specifications and tolerances. They must understand both how the loudspeaker OEMS are testing as part of their incoming QC and also how to implement their own end-of-line measurements to ensure correlation between the two.

Authors: Steve Temme, Listen, Inc. and Viktor Dobos, Harman/Becker Automotive Systems Kft.
Presented at ISEAT 2017, Shenzhen, China

Full Paper

Challenges of IoT Smart Speaker Testing

Quantitatively measuring the audio characteristics of IoT (Internet of Things) smart speakers presents several novel challenges. We discuss overcoming the practical challenges of testing such devices and demonstrate how to measure frequency response, distortion, and other common audio characteristics. In order to make these measurements, several measurement techniques and algorithms are presented that allow us to move past the practical difficulties presented by this class of emerging audio devices. We discuss test equipment requirements, selection of test signals and especially overcoming the challenges around injecting and extracting test signals from the device.

Authors: Glenn Hess (Indy Acoustic Research) and Daniel Knighten (Listen, Inc.)
Presented at the 143rd AES Conference, New York 2017

Full Paper

Evaluation of Audio Test Methods and Measurements for End-of-Line Automotive Loudspeaker Quality Control

In order to minimize costly warranty repairs, automotive manufacturers impose tight specifications and a “total quality” requirement on their part suppliers. At the same time, they also require low prices. This makes it important for automotive manufacturers to work with automotive loudspeaker suppliers to define reasonable specifications and tolerances, and to understand both how the loudspeaker manufacturers are testing and also how to implement their own measurements for incoming QC purposes.

Specifying and testing automotive loudspeakers can be tricky since loudspeakers are inherently nonlinear, time variant and affected by their working conditions & environment which can be change dramatically and rapidly in a vehicle. This paper examines the loudspeaker characteristics that can be measured, and discusses common pitfalls and how to avoid them on a loudspeaker production line. Several different audio test methods and measurements for end-of-the-line automotive speaker quality control are evaluated, and the most relevant ones identified. Speed, statistics, and full traceability are also discussed.

Authors: Steve Temme, Listen, Inc. and Viktor Dobos, Harman/Becker Automotive Systems Kft.
Presented at the 142nd AES Convention, Berlin, Germany

Full Paper

In-Vehicle Audio System Distortion Audibility versus Level and Its Impact on Perceived Sound Quality

As in-vehicle audio system output level increases, so too does audio distortion. At what level is distortion audible and how is sound quality perceived as level increases? Binaural recordings of musical excerpts played through the in-vehicle audio system at various volume levels were made in the driver’s position. These were adjusted to equal loudness and played through a low distortion reference headphone. Listeners ranked both distortion audibility and perceived sound quality. The distortion at each volume level was also measured objectively using a commercial audio test system. The correlation between perceived sound quality and objective distortion measurements is discussed.

Authors: Steve Temme, Listen, Inc. and Patrick Dennis, Nissan Technical Center North America, Inc.,
Presented at the 141st AES Convention, Los Angeles, CA 2015

Full Paper

The Correlation Between Distortion Audibility and Listener Preference in Headphones

It is well-known that the frequency response of loudspeakers and headphones has a dramatic impact on sound quality and listener preference, but what role does distortion have on perceived sound quality? To answer this question, five popular headphones with varying degrees of distortion were selected and equalized to the same frequency response. Trained listeners compared them subjectively using music as the test signal, and the distortion of each headphone was measured objectively using a well-known commercial audio test system. The correlation between subjective listener preference and objective distortion measurement is discussed.

Authors: Steve Temme, Sean E. Olive*, Steve Tatarunis, Todd Welti*, and Elisabeth McMullin*            *Harman International
Presented at the 137th AES Conference, Los Angeles 2014

Full Paper

Advances in Impedance Measurement of Loudspeakers and Headphones

Impedance measurement is often the sole electrical measurement in a battery of QC tests on loudspeakers and headphones. Two test methods are commonly used, single channel and dual channel. Dual Channel measurement offers greater accuracy as both the voltage across the speaker (or headphone) and the reference resistor are measured to calculate the impedance. Single Channel measurement methods are more commonly used on the production line because they only require one channel of a stereo soundcard, which leaves the other free for simultaneous acoustic
tests. They are less accurate, however, due to the test methods making assumptions of constant voltage or constant current. In this paper we discuss a novel electrical circuit that offers similar impedance measurement accuracy compared to complex dual channel measurement methods but using just one channel. This is expected to become popular for high throughput production line measurements where only one channel is available as the second channel of the typical soundcard is being used for simultaneous acoustic tests.

Authors: Steve Temme and Tony Scott
Presented at the 135th AES Conference, New York 2013

Full Paper

Measurement of Harmonic Distortion Audibility Using A Simplified Psychoacoustic Model – Updated

A perceptual method is proposed for measuring harmonic distortion audibility. This method is similar to the CLEAR (Cepstral Loudness Enhanced Algorithm for Rub & buzz) algorithm previously proposed by the authors as a means of detecting audible Rub & Buzz which is an extreme type of distortion[1,2]. Both methods are based on the Perceptual Evaluation of Audio Quality (PEAQ) standard[3]. In the present work, in order to estimate the audibility of regular harmonic distortion, additional psychoacoustic variables are added to the CLEAR algorithm. These variables are then combined using an artificial neural network approach to derive a metric that is indicative of the overall audible harmonic distortion. Experimental results on headphones are presented to justify the accuracy of the model.

Authors: Steve Temme, Pascal Brunet and Parastoo Qarabaqi
Presented at the 51st AES Conference, Helsinki, Finland, 2013

Full Paper