A fair point. I never use the dictation, but I often try to use Siri when the phone is out of reach. My Echo understands me when I speak at a conversational level from across the room with the faucet running. If the dictation is spot on and Siri is still bad when speaking directly into the phone, then yeah, that seems pretty damning.

NLP quality is dependent on ASR quality (automated speech recognition, transforming audio into words) because you can't generally get meaning without words.

ASR quality is dependent on audio quality (plus, of course, a bunch of big-data-driven training) because it's harder to get the words right if you can't hear the sound.

Audio quality is dependent on microphone(s) quality and signal processing (higher signal-to-noise ratio).

And getting that to happen at far-field distances of 10-20 feet in noisy environments is not easy.

So... bottom line: the better the microphones, the higher the ceiling is on NLP performance.

Source: worked on Echo / Alexa for almost 3 years.

Does it really make sense splitting ASR and NLP up into distinct concepts? While I understand in the past they have solved different problems, it's my understanding that to get accurate ASR, grammar files or corpus can be provided to improve the recognition. I've also seen things like Siri autocorrect what is being transcribed as the sequence of words evolves.

Another example is how as an Alexa skill maker, you have to provide utterance / intent mappings. Is that just used to accurately classify intent / entities? Or is it also used to identify which skill to pass a user utterance to as a part of the alexa skill service because there could be very little variation between skill names or inquiries amazon is supposed to actually fulfill when a person is talking ??

> Does it really make sense splitting ASR and NLP up into distinct concepts? While I understand in the past they have solved different problems, it's my understanding that to get accurate ASR, grammar files or corpus can be provided to improve the recognition.

Yes, you are right, in that there are ways to blend the two (use data from one to improve the other). However, in the end of the day, the better the system can determine which words were spoken (using whatever technology), the better it can determine the meaning and the intent, and then decide what to do about it...

> Another example is how as an Alexa skill maker, you have to provide utterance / intent mappings. Is that just used to accurately classify intent / entities? Or is it also used to identify which skill to pass a user utterance to as a part of the alexa skill service

It is primarily used for the former (classify utterance / intent mappings for bootstrapping). Over time, with ML, the goal would be to help understand which "skills" apply to which intents. Unfortunately, today, that's not the case (in Alexa) and that's why the skill-specific keywords or names are still needed. From the user's point of view, it would be preferable to be able to say "Alexa, I need a ride to the airport" and have Alexa figure out whether Uber, Lyft, or the light rail service with a station two blocks from your house is the "best" option for you right now, based on price, availability, and your explicit and implicit preferences. Of course, the system would also have to allow you to specifically request a Lyft, if that's what you want.

(And, of course, it should ultimately be proactive and just offer to get you to the airport when it sees a flight in your calendar, of from having scanned the flight purchase confirmation in your email...)