Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Launch HN: Deepgram (YC W16) – Scalable Speech API for Businesses

news.ycombinator.com

Post reply on HN
Print view View original

21–27 of 27 posts

Re: Launch HN: Deepgram (YC W16) – Scalable Speech API for Businesses

#21

Post by metildaa »

stephensonsco wrote: post #19
dumbfoundded wrote: post #14

Earlier quoted context omitted.

I guess maybe a better way to ask is which acoustic environments do you excel in? In terms of gathering data, I'm curious how to plan to get the 15K audio hours it takes to train each of these models. The most you want to segment it (like through acoustic environment or genders), the more data you need. Do you have a cheap way of generating high quality data?

If you're training from scratch around 10k hours is needed to get a good model, but when you are transfer learning you don't need nearly that much (100 hours gets you a lot). We excel in phone call and meetings settings. I.e. the typical sales/office/support environment.

Baidu trained their DeepSpeech model with 6000 hours of English to get a model similarly accurate to Google/Microsoft, it may just be the type of quick model your using that needs 10k hours to achieve good results.

Mozilla's DeepSpeech is quite interesting, languages like Turkish can get a decently usable (~20% WER) model with just 80hrs of training data (no transfer learning, starting from a clean slate).

Re: Launch HN: Deepgram (YC W16) – Scalable Speech API for Businesses

#22

Post by metildaa »

stephensonsco wrote: post #18
Donald wrote: post #16

Earlier quoted context omitted.

What about custom models?

There's no additional charge for training a custom model when your usage is a minimum of 10k hrs/mo.

So basically a $10k monthly commit is required to train a custom model? Would it be possible to pay for the training itself if you are a lower volume user?

Re: Launch HN: Deepgram (YC W16) – Scalable Speech API for Businesses

#23

Post by stephensonsco »

stephensonsco wrote: post #18
Donald wrote: post #16

Earlier quoted context omitted.

What about custom models?

There's no additional charge for training a custom model when your usage is a minimum of 10k hrs/mo.

You can pay to get training at a lower usage amount too.

Re: Launch HN: Deepgram (YC W16) – Scalable Speech API for Businesses

#24

Post by stephensonsco »

metildaa wrote: post #21
stephensonsco wrote: post #19

Earlier quoted context omitted.

If you're training from scratch around 10k hours is needed to get a good model, but when you are transfer learning you don't need nearly that much (100 hours gets you a lot). We excel in phone call and meetings settings. I.e. the typical sales/office/support environment.

Baidu trained their DeepSpeech model with 6000 hours of English to get a model similarly accurate to Google/Microsoft, it may just be the type of quick model your using that needs 10k hours to achieve good results. Mozilla's DeepSpeech is quite interesting, languages like Turkish can get a decently usable (~20% WER) model with just 80hrs of training data (no transfer learning, starting from a clean slate).

Yep, all good points. One thing to consider is that generalization is a big problem. It's easy to get good on a specific dataset nowadays (like 5-10% word error rate level on academic datasets), but that same model might do 40% WER on data in the wild.

Re: Launch HN: Deepgram (YC W16) – Scalable Speech API for Businesses

#25

Post by stephensonsco »

dumbfoundded wrote: post #14
stephensonsco wrote: post #8

Earlier quoted context omitted.

It's a metric that's hard to nail down because there is so much parameter space that you are flattening into one number. Also it doesn't address the "I care about these five high value words (that are made up), can you recognize them?" like product names and company names. There's ~4 types of audio: Phone call - close microphone - conversational - low bandwidth audio - two way conversation - more industry specific te…

I guess maybe a better way to ask is which acoustic environments do you excel in? In terms of gathering data, I'm curious how to plan to get the 15K audio hours it takes to train each of these models. The most you want to segment it (like through acoustic environment or genders), the more data you need. Do you have a cheap way of generating high quality data?

I didn't answer "Do you have a cheap way of generating high quality data?". We have good ways to do it. They're not that cheap though. It's expensive (organizationally and real $$$) to label large amounts of data no matter what.

But we do utilize our capabilities to better tackle the wild data gathering and labeling. For instance, "is every labeled minute just as valuable as any other?". Definitely not. So if you can find and select only the data you want to label, rather than indiscriminately labeling a bunch, then you can increase your overall efficacy.

Re: Launch HN: Deepgram (YC W16) – Scalable Speech API for Businesses

#26

Post by btown »

stephensonsco wrote: post #5
btown wrote: post #4

(FYI your https://deepgram.com/v2/docs links are giving "error": "Not Found" JSON responses.) I love progress in this space. Something I also think is necessary, though, is innovation in the discoverability interfaces around speech data. Can you search over potential transcriptions weighted by their likelihood, rather than just doing full-text search on the most-likely transcriptions? Can you visualize multiple poten…

This is a seriously fertile area where you get to "define the new interface". It's a big problem though, since few buyers know they want those things. Around 95% of customers come into it with "give me the transcripts" and discover over time they want these other things too (some graphical, some technical). They just didn't know it was available. New GUIs and data representations is a big part of it. Getting accuracy…

Tableau (and the general business analytics space) have done a good job at reframing the problem as: "don't think about what you want as a leader at a company; instead, democratize data access so your team can decide what it wants, and pay for democratization not for your own features." See for instance: https://www.forbes.com/sites/briansolomon/2016/05/04/how-tab...

Arguably Elastic is a success story about bridging the worlds of an API-first technical stack with a democratized non-technical analytics framework. And they started by just powering excellent search, and building value-add layers over time. But they built into a then-vacuum of API offerings, whereas there are many other (potentially inferior, but well-funded) speech-to-text APIs. I'll be avidly following you guys as you navigate the space, and hopefully you're able to find some good "hooks" or uniquely-easy-to-roll-out integration stories that strike a balance between focus on technical excellence and driving awareness in a super-linear way.

Post reply on HN