Slide 1: The Future of Voice Interfaces
The Future of Voice Interfaces Tech That’s Turning Sci-Fi into Reality Eran Belaish, Product Marketing Manager, Audio, Voice and Sensing Robert Schrager, Director of Sales and Marketing at Alango Technologies
Slide 2: CEVA By Numbers
CEVA By Numbers 3X >8 Billion 1 in 3 the market share in DSP over CEVA-powered devices handsets worldwide are any other DSP IP vendor shipped worldwide to date powered by CEVA DSP Audio DSPs 100 >30 >5 billion licensees of Wi-Fi & Bluetooth licensees for imaging and DSP cores in audio/voice IP – and
Slide 3: We Power the World's Leading OEMs
We Power the World’s Leading OEMs Audio DSPs U.S. Flagship Smartphone Audio DSPs Audio DSPs U.S. Flagship Smart Speaker CEVA Proprietary Information 3
Slide 4: Overview
Overview Voice Interfaces - Where are we at Today? Under the Hood of Smart Speakers - Always-on tech, DNN and more Far Field Voice Pickup The Future of Voice Interfaces I Want to Voice-Enable my Product, What’s Next? Q&A CEVA Proprietary Information 4
Slide 5: Voice Interfaces - Where are we at Today?
Voice Interfaces - Where are we at Today? Flagship smartphones by Apple and Samsung are always-listening Most smartwatches offer voice activation Amazon Echo has ignited the far-field conversational assistant trend GoPro Hero 5 can be operated using voice commands Voice activated car infotainment sy
Slide 6: What About Always-Listening?
What About Always-Listening? Leading examples of always-listening devices Flagship smartphones by Apple and Samsung Alexa open Amazon Echo and Google Home the door But, most devices are not always-listening Alexa open Portable speakers (e.g. Amazon Echo Tap) the door! Most smartphones and smartwatch
Slide 7: Under the Hood of Smart Speakers
Under the Hood of Smart Speakers CEVA Proprietary Information 7
Slide 8: Smart Speaker Block Diagram - Main Hardware
Smart Speaker Block Diagram - Main Hardware Far-Field Application Microphone Smart Voice Processor Array Processor (Alexa Client) Bluetooth Audio Speaker & Speaker Processor Wi-Fi Far-Field Voice Pickup is A Main Challenge of Designing a Smart Speaker Solution CEVA Proprietary Information 8
Slide 9: Smart Speaker on a Chip - BT/Wi-Fi/Audio/Voice/Sensing
Smart Speaker on a Chip - BT/Wi-Fi/Audio/Voice/Sensing Smart Bluetooth and Audio and Application Speaker Wi-Fi Solution Sensing Solution Processor Chipset HW Audio/Voice HW VAD D/A PDM2PCM DSP Input Processing SW Output Processing SW • Beamforming • SBC/AAC/MP3 codec • AEC & Dereverberation • Audio
Slide 10: Introduction to Always-on Technology by CEVA
Introduction to Always-on Technology by CEVA CEVA Proprietary Information 10
Slide 11: Always-on Technology by CEVA
Always-on Technology by CEVA Market Leadership Always-on Development Kit Smartphone flagships are always- listening and started a strong trend Power is the main consideration by far CEVA offers the world’s lowest power solution, already in mass production CEVA powers Voice Trigger in Join Our Link
Slide 12: From Ultra Low Power to High Performance
From Ultra Low Power to High Performance Ultra low power High Performance Battery operated devices Smart home devices, often AC (smartwatches, smartphones etc.) powered (digital assistants etc.) Ultra low power e.g. near field High performance e.g. far field voice always-on voice activation activati
Slide 13: Ultra Low Power or Extra Horse Power?
Ultra Low Power or Extra Horse Power? CEVA-TL410 Ultra Low Power DSP CEVA-X2 High Performance DSP CEVA is powering the world’s CEVA -X2 is a high performance hub lowest power voice activation chips for sensing and connectivity workloads CEVA -TL410 can be always-on with Can handle multiple tasks fro
Slide 14: Deep Neural Networks (DNN)
Deep Neural Networks (DNN) CEVA Proprietary Information 14
Slide 15: DNNs Have Enabled a Breakthrough
DNNs Have Enabled a Breakthrough ► All DNNs employ similar concept ► Deep learning - offline training with massive data sets ► Use the generated DNN to classify/filter real time signals ► DNNs on CEVA platforms ► Speech recognition – Sensory is in mass production ► Noise reduction – voice isolation
Slide 16: DNN Voice Isolation demo
DNN Voice Isolation demo Cafeteria Pub Raw Raw Filtered Filtered ► The DNN can tell a human voice Competing Speakers from other sound sources and isolate it Raw Filtered CEVA Proprietary Information 16
Slide 17: Far Field Voice Pickup
Far Field Voice Pickup CEVA Proprietary Information 17
Slide 18: Far-Field Voice Pickup Processing
Far-Field Voice Pickup Processing VAD Input Beamforming AEC Voice Trigger Output Ultra low power Handle Multiple Multi-mic Stereo Acoustic Voice trigger Output recorded hardwired Voice Microphones Beamforming Echo Canceller based on Deep voice (encoded Activity including Neural Networks using Opus)
Slide 19: Preprocessing of Far-Field Speech with Alango's VEP
Preprocessing of Far-Field Speech with Alango’s Voice Enhancement Package (VEP) for Improving Speech Recognition Rate CEVA DSP + Alango VEP CEVA Proprietary Information 19
Slide 20: Challenges of Far-Field Voice Pick-Up
Challenges of Far-Field Voice Pick-Up • Noise • Acoustic Echo • Reverberation • User distance NOISE SPEECH ECHO Wall Reverberant Speech CEVA Proprietary Information 20
Slide 21: Voice Enhancement Package (VEP)
Voice Enhancement Package (VEP) CEVA Proprietary Information 21
Slide 22: Voice Enhancement Package (VEP)
Voice Enhancement Package (VEP) Other microphone configurations are possible. Please consult with Alango on selecting the most optimal configuration for a specific acoustic design. • VEP processing uses N signals of N separate microphone sensors, cleans them of acoustic echo, and forms M acoustic be
Slide 23: Implementation of 4mic VEP
Implementation of 4mic VEP “4mic” VEP with 8 beams NS NS Automatic Speech AEC NS Recognition System AEC (ASR) NS N-Mic or NS Beamformer AEC Keyword Recognition NS (KWR) AEC NS NS M=8 N=4 8 beams 4 mics CEVA Proprietary Information 23 23
Slide 24: Test Results: 2~4mic VEP
Test Results: 2~4mic VEP Tested Voice Triger recognition (hit rate) with Alango multi-directional beamforming in realistic conditions for different users, noise levels and distances Test: Test setup and components: Compare voice trigger recognition rate in office • Real meeting room meeting room, us
Slide 25: Test Results: Test setup configuration, top view
Test Results: 2~4mic VEP Test setup configuration, top view (4-mic smartphone form factor) Recording and collecting data on PC Speaker (turned towards the wall) Produces 75 dB SPL noise level at DUT location 1.6 m Mouth simulator 0.5 m 4m DUT Table Distances from 0.5 to 4.0 meters away from DUT Larg
Slide 26: Test Results: 2~4mic VEP
Test Results: 2~4mic VEP CEVA Proprietary Information 26 26
Slide 27: VEP Status
VEP Status VEP porting and optimization on CEVA DSP (running on CEVA demo, 6 mics) May 2017 Demo voice activated media device shown at CES2017 (running on PC) Jan 2017 VEP algorithm and software development 2015~2016 CEVA Proprietary Information 27 27
Slide 28: Going Multi-Mic
Going Multi-Mic ► Configurable number of mics from 1 to all 13 including popular setups of 4, 7 and 8 mics ► A full set of drivers and SW infrastructure is mandatory Voice Noise Acoustic echo Far-Field Voice Pickup Requires Abundant DSP Horsepower and A Multi-microphone HW Setup and SW Infrastructur
Slide 29: The Future of Voice Interfaces
The Future of Voice Interfaces CEVA Proprietary Information 29
Slide 30: Near Future of Voice Interfaces
Near Future of Voice Interfaces Voice controlled everything Conversational personal assistants everywhere Mobile, wearable, smart home, smart car, medical Mostly driven by search and e-commerce giants Consumer IoT switches to voice-first UI Home robots, thermostats, AC, white goods Smart mics can vo
Slide 31: The Fascinating Future of Voice Interfaces
The Fascinating Future of Voice Interfaces When People and Machines Talk… CEVA Proprietary Information 31
Slide 32: The Fascinating Future of Voice Interfaces
The Fascinating Future of Voice Interfaces Use natural language to control any device Voice trigger will no longer be mandatory Alexa, what’s the weather in ummm, oh I “Alexa” forgot…San Diego! “OK Google” Optional unlike “Hi Siri” today’s systems “Hey Cortana” Integration with computer vision Impro
Slide 33: The Fascinating Future of Voice Interfaces
The Fascinating Future of Voice Interfaces Human-like contextual awareness Always-on voice authentication Human-like memory Emotion detection Human-like Voice Interaction with Every Machine CEVA Proprietary Information 33
Slide 34: I Want to Voice-Enable my Product, What's Next?
I Want to Voice-Enable my Product What’s Next? CEVA Proprietary Information 34
Slide 35: CEVA is Powering the Voice-First Revolution
CEVA is Powering the Voice-First Revolution Samsung Galaxy S7 DSPG’s DBMD4 with CEVA Inside Voice-Enabling Your Product? Tell us About it Our Customers Have Shipped Over 5 Billion Voice-Enabled Devices CEVA Proprietary Information 35
Slide 36: CEVA Sensing/Audio/Voice Ecosystem
CEVA Sensing/Audio/Voice Ecosystem Category Partner Offering Category Partner Offering Life Vibes Voice Experience DRA Codecs Noise Voice Comm. Package Real Audio Reduction / Flexbeam TruMedia HD/StudioSound 3D Echo Cancellation Noise reduction Speaker Correction Noise reduction ZIRENE Sound/3D Trul
Slide 37: More than 150 Audio/Voice SW Modules
More than 150 Audio/Voice SW Modules Noise Reduction Always-on & Post- Dolby HD- RTOS Voice & Echo Audio DTS HD-Audio NUI processing Audio Cancellation Sensory Sensory FreeRTOS; G.723; NXP SW Life MP3; DTS/SRS Dolby TrueHD; Master Audio; TrulyHandsFree; TrulyHandsFree; Express Logic G.729; Vibes Voi
Slide 38: DSP Library
DSP Library Fourier Filter Functions Vector Operations Math Functions Transformations Autocorrelation Bit-Reverse Permutation Addition Division Cross-correlation Complex FFT Subtraction Square Root Convolution (Block FIR) Real FFT Dot Product Inverse Square Root Block LMS Filter Complex Inverse FFT
Slide 39: Architecture Block Diagram
Architecture Block Diagram Feature Configuration Pipeline 10 VLIW 5 way SIMD [bit] 64 Scalar Units 2 MAC [16x16-bit] 4 MAC [32x32-bit] 2 SP Floating-Point Optional Data Memory width [bit] 128 Branch Target Buffer Optional Data Cache Optional Instruction Cache Optional A Unified DSP for Intensive Mul
Slide 40: CEVA-ToolBox Software Development Environment
CEVA-ToolBox™ Software Development Environment Highly focused on Profiler – Function Graph two main objectives: Performance of generated code, especially cycle count and code size User experience, with Profiler – Cache Performance emphasize on ease of Eclipse IDE / programming, user Debugger View in
Slide 41: Audio/Voice/Sensing Value Proposition
Audio/Voice/Sensing Value Proposition Advance Sound Processing Vanilla or Custom Made Solutions ► CEVA-X2: A powerful DSP core ► Extensible architecture Dual 32-bit MAC, Quad 16-bit MAC, 128-bit Customer differentiation through application data bandwidth lower MCPS specific ISA Superb control capa
Slide 42: Thank You! Questions?
Thank You! Questions?