Microphone Integration – Can Improve ARS Accuracy? Tom Houy
2 Telematics\ Infotainment Solutions
3 One Alternative Wind noise reduction Anti-echo boom
4 Improving Voice Quality Mic positioned a long way away from users mouth High power speaker positioned close to mic
5 Headsets Have Come a Long Way
6 Noise Considerations ASR Impacts: Dialog Design Grammars Discreet / Continuous Word / Phrase Non-Impacts: Languages Processing Reqts Memory
7 Hands Free Considerations Acoustic Echo Cancellation Effect on far end speech recognition Echo Return Loss Capability to handle speaker distortion Ability to adapt to changing Acoustic environment Varied Acoustic Coupling Beating with other echo cancellers Double Talk Barge In
8 Hands Free Considerations Noise Suppression SNR improvement Distortion Interaction with vocoder Voice Quality Time to Converge Noise types Ability to adapt to acoustic environmental factors Automatic Gain Control Microphone Interaction
9 Algorithms Must Mimic Human Processing 1)Sound waves cause the ear drum to vibrate in the middle ear 2)Vibration in the middle ear is translated into fluid motion in the cochlea 3)The hair cells in the inner ear vibrate in a unique pattern corresponding to the frequency spectrum characteristic of the incoming sound 4)The motion of each hair is converted to electrical pulses which are transmitted to the brain 5)The brain separates speech from other sound components for analysis How humans hear How ASR engines hear A/D Component Extraction Voice Extraction 1)Sound pressure is converted to electrical waves by the microphone 2)Electrical waves are converted to a digital representation (numbers) by the analog to digital converter 3)The numbers are processed into frequency components which are analogous to the output of the cochlea 4)The independent frequency components are subjected to temporal analysis and correlated to distinguish speech from noise 5)Speech is extracted from the noisy signal for subsequent processing Independent Component Analysis
10 Algorithms Enable Superior Noise Suppression Component Extraction Voice Component Analysis Voice Extraction Synthesis Voice Extraction The voice extractor modifies each frequency component according to the voice and noise model. The Voice Extractor takes advantage of psychoacoustics principals to minimize noise floor artifacts and perceived voice distortion Voice Analysis The voice analysis block considers temporal and correlative properties of speech and noise to develop a predictive model of the speech components Component Extraction The audio signal is converted to the frequency domain and partitioned according to critical voice components Synthesis The extracted voice components are recombined and converted into the time domain
11 Subband Adaptive Filter F 3 Residual Echo Canceller Programmable Non-Linear Processing Near End Tx Rx D/A A/D F3 Residual Echo Cancellation Uses proprietary estimation methods to improve Echo Return Loss introducing speech degradation Subband Adaptive Filter Uses CSR Detroit’s Proprietary Restored Estimator Reutilization method to quickly model the linear transfer function between the reference signal and the microphone. Allows for rapid and more accurate plant estimation Algorithms create a new approach to Acoustic Echo Cancellation Environment Restoration Algorithms Enable Natural Conversation
12 One Alternative Wind noise reduction Anti-echo boom
13 But the real mass market is still in the future…. What will make headsets attractive to everyone for the auto? Work with any phone Batteries last 8 hours (longer for stereo) Looks nice Great sound quality (at least as good as the phone) Natural to use Not too expensive Noise Suppression
14 DSPs enhance Bluetooth Bluetooth Multimedia devices were launched more than 2 years ago and has more than 10MU Next step: More powerful Bluetooth Multimedia; realising the vision
15 Name announcing – improving safety John Smith calling Eyes stay on the road With name announcing Bluetooth headsets can contribute further to road safety