DESIGNING OPTIMIZED MICROPHONE BEAMFORMERS

Similar documents
THE EFFECT OF PERFORMANCE STAGES ON SUBWOOFER POLAR AND FREQUENCY RESPONSES

Application Notes on the ClearOne Beamforming Microphone Array

VTX V25-II Preset Guide

AMEK SYSTEM 9098 DUAL MIC AMPLIFIER (DMA) by RUPERT NEVE the Designer

LIVE SOUND SUBWOOFER DR. ADAM J. HILL COLLEGE OF ENGINEERING & TECHNOLOGY, UNIVERSITY OF DERBY, UK GAND CONCERT SOUND, CHICAGO, USA 20 OCTOBER 2017

DIGITAL SPEAKER MANAGEMENT UK

Generating the Noise Field for Ambient Noise Rejection Tests Application Note

Natural Radio. News, Comments and Letters About Natural Radio January 2003 Copyright 2003 by Mark S. Karney

Acoustic Camera 2018

MAD A-Series...Flat Panel Surface Planar Arrays

What is the minimum sound pressure level iphone or ipad can measure? What is the maximum sound pressure level iphone or ipad can measure?

Put your sound where it belongs: Numerical optimization of sound systems. Stefan Feistel, Bruce C. Olson, Ana M. Jaramillo AFMG Technologies GmbH

HEAD. HEAD VISOR (Code 7500ff) Overview. Features. System for online localization of sound sources in real time

A Straightforward One-Seat Stereo Tuning Process and Some Notes About Why it Works

Practical Application of the Phased-Array Technology with Paint-Brush Evaluation for Seamless-Tube Testing

PRACTICAL APPLICATION OF THE PHASED-ARRAY TECHNOLOGY WITH PAINT-BRUSH EVALUATION FOR SEAMLESS-TUBE TESTING

Technical Guide. Installed Sound. Loudspeaker Solutions for Worship Spaces. TA-4 Version 1.2 April, Why loudspeakers at all?

How to Speak Fiber Geek Article 4: Single-Mode Optical Fiber Geometries

White Paper JBL s LSR Principle, RMC (Room Mode Correction) and the Monitoring Environment by John Eargle. Introduction and Background:

THE DIGITAL DELAY ADVANTAGE A guide to using Digital Delays. Synchronize loudspeakers Eliminate comb filter distortion Align acoustic image.

Calibration of auralisation presentations through loudspeakers

POSITIONING SUBWOOFERS

OPTICAL POWER METER WITH SMART DETECTOR HEAD

Using the new psychoacoustic tonality analyses Tonality (Hearing Model) 1

Acoustic Echo Canceling: Echo Equality Index

DISPLAY WEEK 2015 REVIEW AND METROLOGY ISSUE

Measurement of overtone frequencies of a toy piano and perception of its pitch

The Cocktail Party Effect. Binaural Masking. The Precedence Effect. Music 175: Time and Space

A Performance Ranking of. DBK Associates and Labs Bloomington, IN (AES Paper Given Nov. 2010)

White Paper Measuring and Optimizing Sound Systems: An introduction to JBL Smaart

Precision DeEsser Users Guide

MULTISIM DEMO 9.5: 60 HZ ACTIVE NOTCH FILTER

Performing a Sound Level Measurement

Standard Definition. Commercial File Delivery. Technical Specifications

Panaray 802 Series III TECHNICAL DATA SHEET. loudspeaker. Key Features. Product Overview. Technical Specifications

Piotr KLECZKOWSKI, Magdalena PLEWA, Grzegorz PYDA

Eventide Inc. One Alsan Way Little Ferry, NJ

DH400. Digital Phone Hybrid. The most advanced Digital Hybrid with DSP echo canceller and VQR technology.

What s New in Raven May 2006 This document briefly summarizes the new features that have been added to Raven since the release of Raven

Binaural Measurement, Analysis and Playback

A BEM STUDY ON THE EFFECT OF SOURCE-RECEIVER PATH ROUTE AND LENGTH ON ATTENUATION OF DIRECT SOUND AND FLOOR REFLECTION WITHIN A CHAMBER ORCHESTRA

An Introduction to TrueSource

The Trinity Church Videos: An Audio Analysis

RoomMatch RM and RM TECHNICAL DATA SHEET. asymmetrical array modules. Key Features. Product Overview. Technical Specifications

Tempo Estimation and Manipulation

Directional microphone array system

Calibrating attenuators using the 9640A RF Reference

LIVERPOOL TLX43. Custom-Engineered Drivers

A Real Word Case Study E- Trap by Bag End Ovasen Studios, New York City

A Real Word Case Study E- Trap by Bag End Ovasen Studios, New York City

In addition, the choice of crossover frequencies has been expanded to include the range from 40 Hz to 220 Hz in 10 Hz increments.

Digital Signal Processing Detailed Course Outline

MultiMac SM. Eddy Current Instrument for Encircling Coil, Sector and Rotary Probe Testing of Tube, Bar, & Wire

M44 Multisonic Imager M86 Multisonic Imager. Operation Manual. Miles Technology Inc. Niles, Michigan USA

DSP Monitoring Systems. dsp GLM. AutoCal TM

ECHOFREE EF1210 MULTI-CHANNEL ACOUSTIC ECHO AND NOISE CANCELLER USER MANUAL

DM8000. # Designed and engineered in the U.K. Advanced Digital Audio Processor

SREV1 Sampling Guide. An Introduction to Impulse-response Sampling with the SREV1 Sampling Reverberator

NOTICE. The information contained in this document is subject to change without notice.

FLOW INDUCED NOISE REDUCTION TECHNIQUES FOR MICROPHONES IN LOW SPEED WIND TUNNELS

The 3-way full range TKT105 is a 600 Watt 10" loudspeaker system

Psychoacoustic Evaluation of Fan Noise

NuQ102 ## ## for FOH, sidefill and delay line loudspeakers, and is an exceptional choice for wedge monitoring applications.

Rx antennas at IV3PRK: the 4-Square Rx Vertical Array

Building Technology and Architectural Design. Program 9nd lecture Case studies Room Acoustics Case studies Room Acoustics

BeoVision Televisions

MultiMac. Eddy Current Instrument for Encircling Coil, Sector and Rotary Probe Testing of Tube, Bar, & Wire

Linrad On-Screen Controls K1JT

LX20 OPERATORS MANUAL

Interface Practices Subcommittee SCTE STANDARD SCTE Measurement Procedure for Noise Power Ratio

Mr. Chris Cocallas University Architect and Director Capital Planning and Construction Colorado School of Mines th St. Golden, Colorado 80401

Using the BHM binaural head microphone

Lab 1 Introduction to the Software Development Environment and Signal Sampling

ECE 4220 Real Time Embedded Systems Final Project Spectrum Analyzer

R-1580A Microwave Downconverter. Product Brochure

Small Room Solutions. High-Quality, Cost-Effective Cinema Audio Products for Small Room Applications

456 SOLID STATE ANALOGUE TAPE + A80 RECORDER MODELS

Area-Efficient Decimation Filter with 50/60 Hz Power-Line Noise Suppression for ΔΣ A/D Converters

Audio Engineering Society. Convention Paper. Presented at the 141st Convention 2016 September 29 October 2 Los Angeles, USA

The Imaging Characteristics of an Array with. L. Kogan 1. January 28, Abstract

Effectively Managing Sound in Museum Exhibits. by Steve Haas

Clock Jitter Cancelation in Coherent Data Converter Testing

JBL f s New Differential Drive Transducers for VerTec Subwoofer Applications:

TOUR SERIES LOUDSPEAKER SYSTEMS

AMS 6ICT. Weather-Resistant Enclosure

What really changes with Category 6

Modula SDI Digital Video Enclosures

RECORDING AND REPRODUCING CONCERT HALL ACOUSTICS FOR SUBJECTIVE EVALUATION

Vortex / VSX TM 8000 Integration

RoomMatch Utility RMU208 TECHNICAL DATA SHEET. small-format foreground/fill loudspeaker. Key Features. Technical Specifications

M R X

VLA-VLBA Interference Memo No. 15

DUBLIN TCX102 ## ## Custom-Engineered Drivers

Academia Sinica, Institute of Astronomy & Astrophysics Hilo Operations

IP Telephony and Some Factors that Influence Speech Quality

AcoustiSoft RPlusD ver

Acoustic Measurements Using Common Computer Accessories: Do Try This at Home. Dale H. Litwhiler, Terrance D. Lovell

Acoustical Testing 1

MULTIMIX 8/4 DIGITAL AUDIO-PROCESSING

ex 800 Series ematrix System

Transcription:

3235 Kifer Rd. Suite 100 Santa Clara, CA 95051 www.dspconcepts.com DESIGNING OPTIMIZED MICROPHONE BEAMFORMERS Our previous paper, Fundamentals of Voice UI, explained the algorithms and processes required for a voice UI system. In this paper, we demonstrate how the different microphone types and array configurations affect performance of voice UI systems, and make specific recommendations engineers and product design teams can use to get the best performance from their voice UI products. As we discussed in the previous paper, beamforming is the processing of signals from multiple omnidirectional microphones to focus in on the sound coming from the direction of the most prominent source (i.e., the user s voice) and disregard sounds coming from other directions. A direction-of-arrival (DOA) algorithm first determines the desired direction of beam focus, then the beamformer algorithm passes the sound from the nearest microphone while manipulating the phase of the signals from the other microphones so sounds coming from outside the beam are reduced in level. Figure 1: Pickup pattern of seven-microphone array similar to the one used in the Amazon Echo. There is one microphone in the center and six microphones evenly spaced around the center mic at a radius of 30mm. The traditional way of evaluating microphone beamformers is to look at their beam patterns. These graphs indicate how the array works as a spatial filter and its ability to enhance desired sounds and remove undesired sounds. The figures below illustrate the performance of a 7-microphone array which is looking towards 90 degrees. The design goal was to have a beam width of 45 degrees. The first plot indicates the array s beam pattern for different frequencies. You will note that the array becomes more directional as frequencies increase from 500 Hz to 4 khz. Then at 8 khz there are unwanted beams which appear at roughly 20 degrees and 160 degrees. This is called spatial aliasing and occurs when the wavelength of the sound is smaller than the spacing between microphones.

Fundamentals of Voice UI page 2. Figure 2: Beamforming performance of a microphone array from a look angle of 90 offaxis. Yellow sections indicate the strongest reception of sound, while dark blue areas indicated the strongest rejection of sound. The graph above shows another way of visualizing the same information. Instead of showing a polar pattern, the spatial information is presented on a two dimensional grid. The Y axis corresponds to the look angle and the X axis corresponds to frequency. The color indicates how much attenuation the array provides in each direction. You ll note that the array is omnidirectional at low frequencies and gets narrower at high frequencies. The unwanted high frequency beams also appear about 5 khz. The question to answer is how does the performance of this microphone array translate into actual performance improves in voice UI. Microphone arrays are designed using sophisticated mathematical algorithms. Most of these algorithms focus on optimizing the beam pattern of the array, with the hope that this will translate into actual improvements in performance. Our experience at DSP Concepts has been that the ratio of the user s voice to the background noise is the key determiner of the performance of the system. This is like a signal-to-noise ratio; the signal is the level of speech and the noise corresponds to other interfering sounds in the room. The performance of wake word algorithms, beamformers, and AECs are all correlated to the signal-to-noise ratio. Our experience at DSP Concepts has been that the ratio of the user s voice to the background noise is the key determiner of the performance of the system.

Fundamentals of Voice UI page 3 Figure 3: Signal-to-noise performance of a typical multiple-microphone array using beamforming So instead of looking at beam patterns, at DSP Concepts we focus on the SNR of an array. For example, if we assume a speech level of 60 db SPL, a background noise level of 50 db, and a microphone SNR of 64 db, then the SNR at the output of the array is shown in Figure 3 above. This setup has an intrinsic SNR of 10 db (the speech is 10 db louder than the background noise) and this is the performance you would achieve with a single microphone. Any SNR improvement above 10 db is due to the microphone array, and the graph shows roughly 6 db improvement above 1 khz. The SNR then starts dropping above 5 khz and this corresponds to the spatial aliasing that we discussed earlier. Evaluating based on SNR provides intuition into how much benefit the array will bring. A 6 db improvement allows you to stand 6 db further away or twice as far as would be possible with a single microphone. As we noted in our previous paper, an improvement of just a few db in signal-to-noise ratio may seem insignificant to many audio professionals, who are accustomed to SNRs that are typically far better than the application demands. For example, improving an amplifier s SNR from -105 db to -109 db will not result in a subjectively appreciable performance improvement. However, SNR is much more important in voice UI applications, where the user s voice may be at the same level as the surrounding noise or the music playback coming from the speaker that houses the voice UI system. Thus, microphone configurations and processing algorithms that can elevate a user s voice just a few additional db above the environmental noise can produce a large improvement in voice-recognition accuracy. Parameters Affecting Array Performance This section presents the SNR performance of various microphone array designs. Parameters evaluated here are: 1) Number of microphones 2) Microphone geometry and spacing 3) Background noise level 4) Microphone noise floor (its SNR) We studied an array in which the microphones were arranged on a circle with a 71mm diameter. Rated SNR of these mics was 64 db. Testing was done in an environment with diffuse-field noise at 50 db SPL, with a speech signal at average 60 db SPL. Beamwidth was 45 degrees, and look angle was 0 except where otherwise specified. Signal processing was performed using DSP Solutions Audio Weaver modular programming package.

Fundamentals of Voice UI page 4 Test #1: Number of Microphones In the chart below, the six-microphone array shows a clear advantage in signal-to-noise ratio at all frequencies from 1000 to 5500 Hz. As the number of mics is reduced, overall SNR suffers, although reducing the number of microphones can actually improve SNR within certain frequency bands. Figure 4: Signal-to-noise ratio of single microphone versus arrays of two to six microphones. The higher the line on the chart, the better the SNR. The polar plots below and on the following page show the pickup patterns of circular arrays using one to six microphones, measured at frequencies of 500, 1000, 2000 and 4000 Hz (thus covering most of the range of human speech). Ideally, the pickup pattern should show a tight beam pointed directly to the right, at the look angle of 0, with little variation at different frequencies. Figure 5: Pickup pattern at four different frequencies for single omnidirectional microphone and two-mic array

Fundamentals of Voice UI page 5 Figure 6: Pickup pattern at four different frequencies for three-, four-, five, and sixmicrophone arrays As can be seen in the polar plots, increasing the number of microphones generally allows for a tighter, more focused pickup beam, but in certain cases adding microphones does not improve performance at all frequencies. For example, three microphones clearly produce a better result at all frequencies than two microphones, however, increasing the microphone count to four improves performance from 500 to 2000 Hz but degrades it at 4000 Hz. The two-, three- and four-microphone arrays produce significant off-axis lobing at 4000 Hz; this reduces system SNR and increases the chance of an inaccurate DOA determination, which could make the beamformer aim in the wrong direction. The two-microphone array, in particular, does a relatively poor job of rejecting sounds from 180. (The three- and four-mic arrays also exhibit this flaw, but only to a significant amount at 4000 Hz.) This error can be especially problematic if the unit is placed near a wall or other large sound-reflecting object, where the reflection might cause the voice UI system to think the user s voice is coming from the wall instead of from the user. The arrays of five and six microphones produce better results, with tightly focused beams on the 0 axis, negligible off-axis lobing, and excellent rejection of sounds from 180.

Fundamentals of Voice UI page 6 Test #2: Microphone SNR Because system SNR is critical to accurate voice recognition, it s tempting to assume that using microphones with higher SNR would improve voice UI performance. To test this assumption, total system SNR was tested with microphones rated at 64 and 70 db SNR, each type arranged in arrays comprising one to six microphones. The main advantage of a high-snr mic would occur at low frequencies, because the improved SNR would permit more aggressive processing of low frequencies, which is where most environmental noise in homes and autos typically occurs. The following two graphs show how microphone signal-tonoise ratio affected the performance of the different microphone arrays, with system SNR shown relative to frequency. The higher the trace is on the chart, the better the SNR and the better the performance of the voice UI system should be. Solid lines show the results with the 64 db SNR microphone; dotted lines show the result with the 70 db SNR microphone. The tests were performed twice: once with 50 db SPL ambient noise, once with 35 db ambient noise. Figure 7: Comparing the system SNR of one- to six-microphone arrays implemented with standard (60 db SNR) mics and low-noise (70 db SNR) mics, measured in a test environment with 50 db SPL ambient noise The graph above shows the results with a 50 db SPL ambient noise field, which is what would be encountered in a typical residential living room with no active sources of direct sound (i.e., no voices, TV, loud appliances, etc.). In this case, the improvement gained by using microphones with better SNR is in most cases barely measurable and would not noticeably improve voice UI performance. The graph on the next page shows the same test conducted in a background noise level of 35 db, which corresponds to a very quiet home environment. Under these conditions, using microphones with better rated SNR has a larger impact, in many cases increasing system SNR by about 1 db. However, note that reducing ambient noise has a much larger impact on system SNR, typically improving it by about 14 db. Thus, the benefits of a 1 db improvement in mic SNR on overall system performance would be insignificant in this case.

Fundamentals of Voice UI page 7 Figure 8: Comparing the system SNR of one- to six-microphone arrays implemented with standard (60 db SNR) mics and low-noise (70 db SNR) mics, measured in a test environment with 35 db SPL ambient noise Test #3: Microphone Gain Matching Like other mechanical devices, microphones exhibit unit-tounit inconsistency. The gain of two samples of the same microphone can vary substantially; a tolerance of ±3 db (for a maximum difference of 6 db in gain between two samples) is common. In arrays of multiple microphones, these inconsistencies might negatively affect system SNR and the overall performance of the voice UI system. Microphones with tighter gain tolerance, or with factory calibration measurements for each mic, are sometimes available, but they are typically more costly. To evaluate the effects of microphone gain mismatch on system SNR, models of theoretical arrays of one to six perfectly matched microphones were tested. Gain mismatches of ±1, ±2 and ±3 db were then introduced into the model, and the tests repeated. The three graphs on the next page show how microphone gain tolerance affected the performance of the different arrays, with system SNR shown relative to frequency. The higher the trace is on the chart, the better the SNR and the better the performance of the voice UI system should be. Solid lines show the results with perfect gain matching; dotted lines show the result with the gain mismatched at ±1 (first chart), ±2 (second chart) or ±3 db (third chart). These charts show that gain mismatches in arrayed microphones can have a large negative impact on system SNR, often comparable to the impact that reducing the number of microphones might have. The effect is particularly noticeable in the bottom chart, with the ±3 db mismatch that is typical of the microphones used in voice UI systems. Compared with using low-noise microphones, reducing ambient noise has a much larger impact on system SNR, typically improving it by about 14 db.

Fundamentals of Voice UI page 8 Figure 9: Comparing the system SNR of one- to six-microphone arrays implemented with mics matched to a tolerance of ±1 db (top), ±2 db (middle) and ±3 db (bottom)

Fundamentals of Voice UI page 9 These tests were performed on a theoretical array without an enclosure. Once the mics are mounted in an enclosure, gain and frequency response will change depending on how and where the mics are mounted and the consistency of the acoustic seals around the mics. For this reason, using mics of better consistency, or supplied with factory calibration data, may not produce an optimal result because the acoustical effects of the enclosure and mounting may introduce performance inconsistencies even with the most tightly matched microphones. The best solution in this case is for the microphone gain to be measured with the mics installed, and the gain for each mic adjusted in software. Ideally, each unit would be individually measured and calibrated in the factory after the product is assembled, so that the software can compensate for any inherent gain mismatch in the mics as well as for mismatches caused by the acoustical effects of the enclosure. Test #4: Microphone spacing Increasing mic spacing in an array might be expected to create greater differences in level among all the mics because the source-to-mic distances will be greater. It will also alter the relative phase among the mics. To find out how spacing affects SNR, arrays using two to six mics were tested, with mics placed on circles ranging from 5 to 71mm in diameter. A three-mic array was tested with the mics placed on circles measuring 40, 80, 160 and 320mm. The following eight graphs show how mic spacing affected array performance, with system SNR shown relative to frequency. The higher the trace is on the chart, the better the system SNR and the better the performance of the voice UI should be. Results for the two- and three-microphone arrays are shown first from an end-fire position (with the source directly in line with one of the mics) and a broadside position (with source equidistant from two microphones). Figure 10: Effect of different microphone spacings on a two-mic array, measured from end-fire (top) and broadside (bottom) positions

Fundamentals of Voice UI page 10 Figure 11: Effect of different microphone spacings on a three-mic array, measured from end-fire (top) and broadside (middle) positions, and from broadside position on a three-mic array with wider mic spacings up to 320mm (bottom)

Fundamentals of Voice UI page 11 Figure 12: Effect of different microphone spacings on four-mic (top), five-mic (middle) and sixmic (bottom) arrays

Fundamentals of Voice UI page 12 To summarize the results from the microphone spacing tests, within the human vocal range: Two-mic array: Best results at 71mm spacing from endfire position, 40mm from broadside. Three-mic array: Best results at 35.5mm from end-fire position and 71mm from broadside position. With wider spacings, increasing spacing to 80mm further improves performance, but greater increases narrow the range of optimum SNR to the point where SNR in some of the vocal range would decrease. Four-mic array: Best results at 71mm. Five-mic array: Best results at 71mm. Six-mic array: Best results at 71mm, although performance at 40mm is not as close to the result at 71mm as it is with the four- and five-mic arrays. Based on these results, placing microphones on a circle measuring 71mm in diameter is generally the best choice with arrays of three to six microphones, if there is sufficient physical space on the voice UI device. With a two-microphone array, results vary considerably depending on whether the source is in line with the microphones or broadside to the mic array. This result suggests that if a two-microphone array is used (either to cut costs or accommodate specific device form factors), the device should be optimized for either end-fire or broadside arrival, and designed so that the user is more likely to address it from the optimal direction. Performance of Existing Products While it is important for a product team to understand how design decisions affect performance of a voice UI system, it is also useful to have benchmarks against which a product can be compared. For this reason, the performance of the mic arrays and algorithms of three existing products were tested. These products are the Amazon Echo and Google Home smart speakers. Also included are measurements of a system using DSP Concepts Voice UI algorithm. Amazon Echo The Echo smart speaker has a top-mounted array of seven microphones one centered with six arranged in an 82mm circle around it, along the top outer edge of the chassis. The following graphs show how well the Echo is able to focus on sound from a certain direction. These measurements were taken at a look angle of 90, perpendicular to the Amazon logo on the front of the unit. The polar pattern below left shows a fairly tight beamwidth averaging about ±30, with very little signal picked up from the rear and only slight side lobing at 8000 Hz. The chart at right shows how the beamwidth varies with frequency. The chart on the next page shows that the system SNR of the Echo s mic array and processing is relatively high, and relatively consistent at most frequencies. Figure 13: Polar pickup pattern (left) and beamforming performance (right) of original Amazon Echo.

Fundamentals of Voice UI page 13 Figure 14: System SNR vs. frequency of Amazon Echo Google Home The Home smart speaker has an array of two microphones, one at each side of its slanted top, spaced 71mm apart. The unit is designed so that a look angle of 0 addresses the two microphones broadside. The following graphs show how well the Home is able to focus on sound from look angles of 0 (broadside) and 90 (end-fire). The chart below left shows the polar pattern of the Echo s mic array from the broadside (0 ) position, which shows strong lobing at 4000 and 8000 Hz, a figure-8 pickup pattern at 2000 Hz and a nearly omnidirectional pickup pattern at lower frequencies. The chart below right shows that the directionality of the array varies greatly with frequency. Figure 15: Polar pickup pattern (left) and beamforming performance (right) of Google Home smartspeaker tested from broadside (180 ) position

Fundamentals of Voice UI page 14 Figure 16: System SNR vs. frequency of Google Home measured from broadside position The chart above shows the system SNR of the Echo s mic array from the broadside (0 ) position. The SNR is optimal only within a narrow range, and is less than that shown in most of the other measurements performed for this paper. The chart below left shows the polar pattern of the Echo s microphone array from the end fire (90 ) position, which shows considerable lobing at 4000 and 8000 Hz, with strong pickup from 270. The chart below right shows that the pickup is strongest from 90 but also relatively strong from other directions as well, indicating that the array s directionality is not well-focused. The chart below right shows the system SNR of the Echo s mic array from the end fire (90 ) position, which is relatively poor and somewhat inconsistent as frequency changes. Figure 17: Polar pickup pattern (left) and beamforming performance (right) of Google Home smartspeaker tested from end-fire (90 ) position

Fundamentals of Voice UI page 15 Figure 18: System SNR vs. frequency of Google Home measured from end-fire position Evaluations of Combined System Performance: Amazon Echo, Google Home and DSP Concepts Voice UI As has been shown, the performance of a voice UI product is a result of many parameters, including form factor, component choice, tuning and software algorithms. This paper will conclude with a comparison of voice command systems on two key aspects of voice UI performance: reliability of trigger word recognition, and acoustic echo cancelling. In both tests, the Amazon Echo and Google Home were compared with a test system from DSP Concepts, employing DSP Concepts Voice UI software package with microphones as specified below, and a UE Roll Bluetooth speaker modified to bypass its internal DSP. Trigger Word Recognition To compare trigger word recognition among the three systems, the Amazon Echo and Google Home were unmodified, and their standard trigger words/phrases ( Alexa and OK Google, respectively) were employed. For the DSP Concepts Voice UI system, a single microphone with advanced noise reduction was used with the Voice UI algorithms, feeding a Sensory Truly Handsfree voice recognition system, which uses the trigger phrase Hello Blue Genie. All units were tested under the same conditions, with multiple nearby speakers producing background babble noise that simulates typical noise and light conversation in a home. The trigger phrases were normalized in level at average 60 db SPL at 1 meter and emitted through a single loudspeaker. The chart below shows the accuracy of trigger word/phrase recognition at different background noise levels. Figure 19: Trigger word recognition accuracy of Amazon Echo, Google Home and DSP Concepts Voice UI tested at different background noise levels

Fundamentals of Voice UI page 16 Once SNR reaches a fairly high level (at least 12 db), the systems all perform well, with Echo exceeding 90% reliability and the Home and DSP Concepts systems exceeding 95%. When SNR drops to 6 db, the DSP Concepts system has a strong advantage, and the Echo s detection percentage climbs above that of the Home. At lower SNRs, performance with all systems falls below 60%, which will likely cause customer dissatisfaction. Acoustic Echo Cancelling (AEC) Acoustic echo cancelling the ability of a voice UI product to reject the sound of the program material coming from its own speaker is critical to achieving accurate voice recognition. To see how overall AEC performance of the three systems compares, the same setup used for the trigger word recognition tests was used, with an XMOS six-mic array replacing the single mic in the DSP Concepts system. Again, all systems were tested in the same environment. IEC shaped noise was played through the speakers to simulate music playback while providing the steady-state stimulus needed for these measurements. The chart below shows the percentage of trigger word recognition at various SNRs, which reflect the ratio of the playback level of the device to the level of the voice. The Echo and Home were used in factory condition with their standard trigger word/phrases ( Alexa and OK Google, respectively); the DSP Concepts system fed a Sensory Truly Handsfree voice recognition system, which uses the trigger phrase Hello Blue Genie. Four different DSP Systems Voice UI algorithms were tested, each requiring a different amount of processing power. Figure 20: Trigger word recognition accuracy of Amazon Echo, Google Home and four DSP Concepts Voice UI algorithms tested at different background noise levels The results show that the Echo s AEC is clearly superior to that of the Home as playback levels increase. The DSP Concepts system exceeds the performance of the Home, and exceeds the performance of the Echo if the more processing-intensive algorithms are used. This result indicates that echo cancelling can be substantially improved if more processing power is made available.

Fundamentals of Voice UI page 17 Conclusions The most important conclusion to draw from this paper is that there is much product development teams can do to optimize the performance of voice recognition systems. By giving the voice recognition AI system the cleanest possible voice signal to work with, engineers can assure the most accurate voice recognition and the greatest customer satisfaction. Here are a few key principles product development teams should keep in mind when designing products that include voice command features: One: Accuracy and reliability of voice recognition are the result of many factors, including form factor of the device, the components chosen, and the algorithms used. Excellent performance in one of these factors does not guarantee reliable performance of the entire system. Two: Generally, the more microphones a product employs for voice pickup, the better. Three to six mics are optimal, although a five-mic array comes close to the maximum possible performance. Three: Microphone spacing of 40 to 80mm is typically best, with 71mm a good all-around choice that can be easily implemented in a wide range of products and applications. Four: Matching the gain of the microphones in an array can substantially improve signal-to-noise ratio. Rather than using microphones with matched or calibrated gain, it is best to match the gain of the microphones at the factory through software, after they are installed in the device. Five: Software-based noise cancelling algorithms can enhance accuracy of voice recognition, but they must be tuned to focus on the frequencies that are most important to voice recognition systems, not to the frequencies most important to the human ear.