Amazon is discontinuing Alexa’s celebrity voices, even if you paid for them | Hacker News Reader

Amazon is discontinuing Alexa’s celebrity voices, even if you paid for them

theverge.com

This is one of the worst run businesses in tech. The fact that they haven't hooked up an LLM to Alexa is insane, even if it's just for some kind of fun magic 8 ball toy thing. My kids already expect it to be able to answer arbitrary questions... now it just fails.

well, it didn't help they fired a good chunk of the alexa team.

"It’s not exactly clear why Amazon has decided to discontinue the feature, but it could be a sign of broader Amazon’s broader Alexa woes."

"Last November, a report from The Wall Street Journal revealed that Amazon CEO Andy Jassy heavily scrutinized the company’s Alexa sector, as its operating losses exceeded $5 million in recent years."

"The company also announced layoffs affecting 18,000 employees earlier this year, with Amazon hardware chief Dave Limp telling CNBC that 2,000 of those cuts affected workers in his division, which is in charge of Alexa and Echo products."

It turns out that the "killer app" for Alexa / Echo devices is being a voice-activated kitchen timer and playing music. Some users also use it as a voice-activated controller for smart home devices, but they are only a small fraction of users. This means that Alexa costs more to operate than it brings in in revenue, so they are looking to cut costs in any way they can.

Hooking up ChatGPT or similarly powered backend seems like a next step to stay competitive, but that will cost even more to operate. Expect to see a smarter AI-powered Alexa available for a subscription that gives better answers and has more skills.

Too bad Alexa failed even at basic tasks, I bought an Amazon BRANDED smart microwave/oven, which worked fine until on Thanksgiving I told it “bake at 400 degrees Fahrenheit for 30 minutes” from the upstairs (my kid had put the food in while I was in the shower), and it wasn’t until 20 minutes later I smelled smoke that I realized Alexa had somehow interpreted “_ at 400 degrees Fahrenheit for 30 minutes” (I looked at the Alexa log of how it had parsed my words, to try to figure out where it went wrong) to mean “please burn down my house”.

I’ve also noticed they aren’t selling that microwave anymore.

Your experience is just like the "Carousel of Progress" at Disneyworld!

It’s funny you say that, because my wife said the exact same thing and laughed so hard when we went to the carousel of progress last year!

> Expect to see a smarter AI-powered Alexa available for a subscription that gives better answers and has more skills.

Umm, no. If I owned these devices (which I don't, because it's spyware), I'd push for the ability to plug in my model of choice (e.g., Alexa, ChatGPT, BARD, my own model).

Because Alexa lives in the Cloud, this should not impact the hardware I don't believe. Firmware updates can happen regardless of my voice-assistant of choice.

Smaller models can be stored locally on device, or accessible via a skill. Larger models can be stored in the cloud, just like Alexa. However, I don't think that a personal voice assistant language model requires cloud. Unless of course Amazon removed the on-device memory. If they did, then that decision essentially rendered 100 million devices e-waste, as they are surveillance bricks without the "subscription service" Amazon will require owners to purchase.

Yikes, what a mess.

How difficult is it now to train your own voice models? I remember seeing years ago that people had Obama, JayZ, and Eminem able to say arbitrary things, the issue was the hardware and having a massive amount of text that matches what they are saying.

The Home Assistant team really seems like they are going to accomplish their goal of fully local voice assistants this year, but are avoiding any discussion of training or using any model that isn't fully licensed and open sourced. Piper seems nice, but there are dozens of other voices I would like instead, including my own voice.

there was a post yesterday on HN, someone made a deepfake kit that can make a video from a single video. It pulls a lot of pretrained data automatically, which I take it is how it accomplishes it.

So maybe with enough demand, there will be a similar voice solution.

This sort of exposes how very little I understand of this technology. Maybe I can Cunningham my way into learning something new very affordably here.

Reply on news.ycombinator.com