(Disclaimer: I'm not an expert in the field and might have misunderstood the current state of affairs. Please correct me if you spot a mistake).
My understanding is that the computational costs of speech-to-text are prohibitive for many consumer startups, leaving them with no choice but to integrate with Alexa, Google Assistant and Siri, who subsidize S2T costs in exchange for near-complete control over the user relationship.
Unfortunately it looks like the economics of voice assistants will drastically favor Big Tech, and make it harder for new entrants to compete.
For reference current pricing for gce speech seems to be around:
$0.006 USD / 15 seconds* (Each request is rounded up to the nearest increment of 15 seconds.)
Do not know how much of speech processing a typical startup would do, so cannot really give an estimate.
I remain a bit baffled by this. Speech to text on a desktop PC with a 400 MHz processor back in the day was "pretty decent". I don't know if I really feel like what people get with Siri, Alexa, Cortana, Google, etc. is actually significantly better... so why is this prohibitively expensive in our modern era of multi-gigahertz phones?
Because customers aren't satisfied with pretty decent when it comes to speech to text, if it isn't state of the art there's someone just around the corner doing it better....
I mean, is the processing power of today's voice software actually impossible to do without a cloud service subscription, or is it that the developers working on the problem today would rather not work on it for local operation because they can't charge cloud pricing or collect user data that way?
Most local voice processing solutions seemed to stop dead in their tracks, developmentally, when companies figured out we could make this a cloud service, but I don't know if I really believe it's technically unfeasible the other way.
Speech-to-text still works when my Pixel 2 is in airplane mode, so there must be some local fallback. It's good, but not as good as the cloud-backed version.
There is at least one startup working on a local solution (I forget their name), but my understanding is that it's still technically inferior - for example it can only support very small vocabularies.
The bottleneck is simply raw gpu compute power available.
I completely agree with you that the technical gap might be caused in part by incentives from cloud companies. Why invest in researching something that makes your core business less valuable?
Amazon and Google both most likely have more to gain from reducing their compute requirements for in house voice recognition than from keeping the cost high for others.
My guess is that the difficulty is consumer behavior, not hardware. Despite how Amazon's success makes it look, getting people to buy an Internet-connected speaker with a microphone and install it in their home is not that easy. It would probably be even harder for a startup to sell them a second one from a different vendor. An app for the device they already have seems like an easier sell?
current state-of-the-art recognizers in the cloud are probably done on CPU systems, so it could be done on-device. once they switch to TPUs than the consumer devices need to get a little chip as well to complete.
16 comments
[ 3.0 ms ] story [ 50.0 ms ] threadMy understanding is that the computational costs of speech-to-text are prohibitive for many consumer startups, leaving them with no choice but to integrate with Alexa, Google Assistant and Siri, who subsidize S2T costs in exchange for near-complete control over the user relationship.
Unfortunately it looks like the economics of voice assistants will drastically favor Big Tech, and make it harder for new entrants to compete.
Do not know how much of speech processing a typical startup would do, so cannot really give an estimate.
Most local voice processing solutions seemed to stop dead in their tracks, developmentally, when companies figured out we could make this a cloud service, but I don't know if I really believe it's technically unfeasible the other way.
The bottleneck is simply raw gpu compute power available.
I completely agree with you that the technical gap might be caused in part by incentives from cloud companies. Why invest in researching something that makes your core business less valuable?