"Edge AI": encapsulating machine learning classifiers in lightweight, energy-efficient, airgapped chips

Writing in Wired, Boing Boing contributor Clive Thompson discusses the rise and rise of "Edge AI" startups that sell lightweight machine-learning classifiers that run on low-powered chips and don't talk to the cloud, meaning that they are privacy respecting and energy efficient. Read the rest

Rhasspy: a free/open voice assistant toolkit that's fully offline

US Air Force research scientist Michael Hansen created Rhasspy as a privacy-oriented alternative to surveilling "voice assistant" products like Google Assistant, Alexa and Siri; the free/open project supports dozens of languages from German, French and English to Mandarin, Vietnamese and Russian, and is designed to run on Raspberry Pi-based devices. Read the rest

2019 was the year of voice assistant privacy dumpster fires

2019 was the "I Told You So" year for privacy advocates and voice assistants: the year in which every company that wanted you to trust them to put an always-on mic in the most intimate places in your home was revealed to have allowed thousands of low-waged contractors to listen in on millions of clips, many of them accidentally recorded: first it was Amazon (and again!), then Google, then Apple, then Microsoft. Read the rest

Like Amazon, Google sends voice assistant recordings to contractors for transcription, including recordings made inadvertently

After Bloomberg revealed that Amazon secretly sent recordings from Alexa to subcontractors all over the world in order to improve its speech-recognition systems, a whistleblower leaked recordings from Google Home to investigative reporters from VRT, revealing that Google, too, was sending audio clips from its voice assistant technology to pieceworkers through the Crowdsource app. Read the rest

The army of contractor-linguists who power Google Assistant say they had their wages stolen

The reason Google Assistant (that's the product you invoke when you say "OK Google" to your device) works reasonably well is that the Pygmalion team -- a small army of linguists -- work long hours handcrafting variations on common phrases ("set a timer for five minutes," "remind me in five minutes," "in five minutes, remind me...") and grammars that allow the system to correctly respond to your queries. Read the rest

New Amazon patent application reveals "solution" to missed Alexa instructions: always on recording

When you talk to Alexa and other voice assistants, you have to phrase your requests by starting with their "wakeword" ("Alexa" "OK Google" "Siri" etc). Read the rest

Amazon stores recordings of Alexa interactions and turns them over to internal staff and outside contractors for review

Bloomberg reporters learned that -- despite public pronouncements to the contrary -- Amazon has an "annotation team" of thousands of people all over the world, charged with reviewing recordings made by Alexa devices in the field, with both staffers and contractors listening to conversations that Alexa owners have had with and near their devices. Read the rest

The Russian equivalent to Alexa is a "good girl" but not too friendly, and is totally OK with wife-beating

Yandex is Russia's answer to Weibo, an everything-under-one-(semi-state-controlled)-roof online service, and its answer to Alexa is Alisa. Read the rest

Voice assistants suck, but they suck worse if you have an "accent"

Research into the shittiness of voice assistants zeroed in on a problem that many people were all-too-aware of: the inability of these devices to recognize "accented" speech ("accented" in quotes because there is no one formally correct English, and the most widely spoken English variants, such as Indian English, fall into this "accented" category). Read the rest

Voice assistants suck (empirically)

New research from legendary usability researchers The Nielsen (previously) Norman (previously) Group finds that voice assistants are basically a hot mess that people only use because they are marginally better than nothing. Read the rest

What Google Assistant calls will actually be like

Perhaps you were slightly unnerved by Silicon Valley cheering Google's startlingly convincing and conversant simulation of a human voice! You know they don't really give a damn about online fakery and abuse, so you know they won't give a damn what ends this tech is put to.

Thankfully, it probably won't work quite so well as the demo. Mr. Bandwagon's edit of Google's presentation is great, an artifact popping in perfect form from the near future's mercifullly unequal distribution.

The thing is this: if humans don't know they're talking to robots, they won't talk in a way robots will understand, which is what we tend to do with Siri and other voice assistants. It'll take a lot of machine learning to grasp the complexities and vagaries of truly natural human speech, a point so obvious that everyone assumes it will obviously be overcome.

Maybe we'll find ourselves talking robotically for the benefit of machines we believe are human. But it's more likely we'll become swiftly inoculated against The Voice, attuned to its little shibboleths and flaws--no Voight-Kampff test necessary--at least for now. We'll just be angrier than ever at our phones, hanging up at the first sign of Robocall 2.0, until it becomes so pervasive we have no choice. Read the rest

EFF tells the Copyright Office: we don't know how to make voice assistants better, but here's how not to make them worse

Every three years, the US Copyright Office asks for proposals for exemptions to Section 1201 of the DMCA, which bans breaking DRM; in 2015, the Electronic Frontier Foundation won a broad "jailbreaking" exemption to modify the firmware of phones and tablets; this year, we're asking for that permission to be extended to smart speakers like Alexa/Echo, Google Home, Apple HomePods, and the smaller players in the market. Read the rest

Adversarial examples: attack can imperceptibly alter any sound (or silence), embedding speech that only voice-assistants will hear

Adversarial examples have torn into the robustness of machine-vision systems: it turns out that changing even a single well-placed pixel can confound otherwise reliable classifiers, and with the right tricks they can be made to reliably misclassify one thing as another or fail to notice an object altogether. But even as vision systems were falling to adversarial examples, audio systems remained stubbornly hard to fool, until now. Read the rest

Voice assistants can be hacked by commanding them with inaudible ultrasonic speech

In DolphinAttack: Inaudible Voice Commands, researchers from Zhejiang University demonstrate an attack on popular voice assistants in laptops and mobile devices from Apple, Google, Amazon, Microsoft, Samsung, and Huawei: by commanding these assistants using speech that has been shifted to ultrasonic ranges, they are able to hijack devices in public places without their owners' knowledge. Read the rest