#258 – An Interview with Bertrand Irrisou and Gerald Friedland of Audeme

Podcast: Play in new window | Download

Subscribe: Apple Podcasts | RSS

Welcome Bertrand Irrisou and Gerald Friedland of Audeme!

• Audeme has a kickstarter for the Movi platform, a speaker independent, cloudless speech detection Arduino shield.
• The Movi also solves a range of privacy issues, as no audio goes to the cloud. This is in contrast to the Samsung TVs sending voices to 3rd party companies.
• Dave has done a video on a voice detection chip from the 80s, the VCP200. There was also a similar version called the SPO256.
  [tube]https://www.youtube.com/watch?v=kFth9K_IvwA[/tube]
• Chris mentions that there is a big difference between a modern speech synth and a Speak and Spell.
• ELIZA was a 1970s NeuroLinguistic Programming product that provided a surprisingly cogent speech engine. 
• All of these chips are based on phonemes, which allow a conversion from sound to actual word meanings.
• Audeme is using the Allwinner A13 chip which has a single core Cortex A8. It also has a built in synthesizer.
• They are running Debian on board, which allows low level driver control and high level software (via Linux).
• Chris has seen something similar on his mobile, the MotoX. It has a coprocessor that listens for audio cues.
• Algorithms are open source, will go online when released.
• Audeme will be at the Embedded System Conference later this month. 
• Users can expect a response time of 0.5 sec normally.
• The audio front end has an Automatic Gain Control op amp, which can implement echo cancel if an external mic array is used.
• The Movi was recently used to control the Romibo robot. It needed to “listen” through a set of fur.
• The expected frequency response of the Movi is 50 – 16 kHz (in contrast to a 300-3kHz response for a telephone)
• The processing uses MFCC. This is in combination with a DCT -> Mel spacing -> DCT cycle.
• Using the Movi with a push To talk (PTT) helps to reduce overall errors. Callsigns (like Siri) also remove false triggering. Chris wants to  talk to it like JARVIS (from Iron Man).
• The board has a peak power requirement of 3W.
• The interface, including for the Arduino, uses a RS232 protocol for the interface. This means it can be used for any platform, not just Arduino. 
• The code requires is mostly C/C++ for low level, Python for glue, Shell scripting for OS stuff. It’s a wide variety of programming languages and platforms.

You should consider backing this project via their Kickstarter page. You can also see a variety of videos of this board in action on their Facebook page.