adrianbg

It should be working in the UK now.

It should be working now.

Should be working now.

Only that that's all I've had time to do so far. What alternative did you have in mind?

Thanks for trying it. I don't see any errors from today on my end. Would you mind emailing me at adrian@patter.io to debug?

India, Canada, the UK, and other English-speaking countries are in Amazon's review process. Should be working in another day or two.

I ranted about this in the unofficial Alexa Slack team and the people there gave some plausible (but unsatisfying) reasons why it works like that. I'm not a huge fan of Amazon though, so I'm happy to blame them. Essentially it boils down to different regions being treated as different "languages" as well as having features rolled out to them at different rates.

Made a new version and submitted it for review. If all goes well, it should be working in a day or two. Let me know how it goes.

I.... don't know. But I doubt it :)

Oh man, def not intentional. I'm Canadian too.

Hmm sorry. Where abouts are you? I didn't intentionally put on any restrictions.

Google's speech recognition is supposed to be much better, though not sure if they'd allow "one intent to rule them all" like you want.

For free-form speech recognition in Alexa, the best option I've seen mentioned on the public Alexa Slack team is using the "SearchQuery" slot. So you'd still have to make a weird catch-all intent that would eat up some of the words (and you wouldn't be able to see them). At the same time, you shouldn't assume that Alexa will give you very good results with such loose constraints. Even in my simple skill it's very bad about confusing certain pairs of words.

I find Amazon's products often have bizarre constraints like that. They seem to do well anyway though :/

Thanks for your comments! I wonder if I can create some kind of personalized briefing based on which stories people seem interested in.

My guess would be that they have separate speech recognition / synthesis models. Sorry, I should try and get this thing out to other countries too.

Yeah, I don't think it's intended to be interactive. More like a really quick curated news update.

It's closed source. Sorry. I've thought about things like trying to crowdsource the summaries, since people sometimes write "tl;dr" comments on posts.

Sometimes I wish it didn't. They don't give you the original audio, any kind of confidence score, or even alternative hypotheses. It's really a pretty rigid platform. A lot of things that seem like they should be reasonable are impossible. Eg., I'd prefer to just say a list of post titles and let people interrupt Alexa when they hear something they like. That is impossible right now without pretty serious hacks.

Yeah.. turns out that's actually a friend of a friend. I'm curious to see what people prefer for obvious reasons :). The advantages of my approach are personalization, interactivity, and scalability beyond just HN. Theirs is probably going to be a nicer experience than I can do for non-interactive use.

Thanks!

Alexa does most of the hard stuff: speech recognition / intent detection, and speech synthesis.

My back end is a simple Python service on GCP that handles HTTP requests from Alexa. The same service also downloads the HN front page from the FireBase mirror and gets summaries from this API:

https://rapidapi.com/textanalysis/api/Text%20Summarization

It's not perfect though, so I may switch to a more expensive summarization API, supplement it with manual summaries, and/or train my own summarization model.

Yeah, RSS feeds are definitely on my mind as well.

SF

Thanks!

Right now I'm using this:

https://rapidapi.com/textanalysis/api/Text%20Summarization

but may switch to a more expensive one and may supplement with manual edits.

Yes they are already penalized. I personally am kinda happy about this blacklisting.

The startup doesn't need to file the H1B themselves. The employee can transfer it from another company.

You don't know what you're talking about. I am an immigrant working at a startup and I know many others.

That would penalize startups, giving only large companies access to the foreign labour pool. It's rare that a problem like this has a simple solution.

I defer to your obvious familiarity with British TV.

Yes, exactly. I do wonder whether a similarly good end-to-end system could be trained by constraining the alignments as I've seen done in some papers.

That's fair. I think both approaches are useful in different situations. Mozilla seems to be focussing on Siri-like use-cases as opposed to dictation. Even for dictation, for many people having to train the system themselves is more work than they're willing to do. I'm sure what you want will exist eventually :)

Yeah. I don't get the impression that the Kaldi core team has been trying very hard recently to get SOTA on eval2k/switchboard. This number uses one acoustic model with a trigram LM decode + fourgram rescoring -- there isn't even a neural net language model in there. If I remember correctly, Microsoft's first "human parity" result used something like three acoustic models and at least four types of language models. This Kaldi model is competitive with the best single acoustic model Microsoft used.