This article is contributed. See the original author and article here.
Welcome back to another blog post about the Azure Percept DK!
In the previous post – Azure Percept – First Look – you’ll remember that we had a first look at the percept and what it was all about. Well, in this post we’ll take a look at the Azure Percept Audio Module, which allows for the recognition of Custom Keywords and Commands (among other things).
What is the Percept Audio?
The Azure Percept Audio (sometimes called the Percept Ear) is a “System on a Module” or SoM, which is designed as the Audio Interface for Audio Processing at the edge for the Azure Percept.
Along with the Carrier Board, Azure Percept Studio, Microsoft LUIS and Speech, the system can recognise keywords and commands to control devices using voice at the edge. This works both online and offline with the aid of the Carrier Board.
Azure Percept Audio Specifications
The basic specs for the Azure Percept Audio SoM are;
- Four-microphone linear array and audio processing via XMOS Codec
- 2x buttons
- 3x LEDs
- Micro USB
- 3.5 mm audio jack
You can find the full specifications here
Who’s it for?
Like the Vision SoM, Microsoft clearly have a set of industries in mind for the Azure Percept Audio SoM;
- Smart Buildings
With applications such as;
- In-room Virtual Concierge
- Vehicle Voice Assistant and Command/Control
- Point of Sale Services and Quality Control
- Warehouse Task Tracking
This becomes clear later when we look at the sample applications we can spin up in a minute.
Azure Percept Audio – Required Services
The Azure Percept Audio SoM makes use of a couple of Azure Services to process Audio;
LUIS (Language Understanding Intelligent Service):
LUIS is an Azure service which allows interaction with applications and devices using natural language.
Using a visual interface, we’re able to train AI models without the need for deep Machine Learning experience of any kind.
The Azure Percept uses LUIS to configure Custom Commands, allowing for a contextualised response to a given command.
Cognitive Speech is an Azure Service offering Text-to-speech, speech-to-text, speech translation and speaker recognition.
Supporting over 92 languages, this service can convert speech to text allowing for interactivity with apps and devices.
On the flip side, with support for over 215 different voices in 60 languages, the Speech Service can also convert Text to-Speech improving accessibility and interaction with devices and applications.
Finally, the Speech Service can also translate between 30 different languages, allowing for real-time translation using a variety of programming languages, which I think is a really cool use case.
The Percept uses this service amongst other things, to configure a wake word for the device, by default this is the word “computer“. (See Star Trek IV – The Voyage Home!).
Azure Percept Audio – Sample Applications
If we navigate to Azure Percept Studio, from the Overview Page we can select the “Demos & tutorials” tab at the top;
If we scroll to the bottom of this page, we have some links to some Speech tutorials and demos.
The first thing we’ll choose is “Try out voice assistant templates”. Clicking this link presents us with a fly out with a selection of templates to choose from;
You can see here a selection of Sample Templates that speak to where Microsoft expect the Percept Audio to be used. All of these a limited to voice commands interacting with an environment. I’ll speak later about some scenarios I’d actually like to be thought about that are outside of this use case.
For now, we’ll deploy one of these pre-backed samples and see how it works!
Azure Percept Audio – Hospitality Sample Template Setup
Choosing the “Hospitality” option, agreeing to the terms and continuing on, we’re shown the resource creation flyout.
Here we can select the subscription and resource group we’d like to deploy the various resources to.
We’re also prompted for an Application Prefix. This allows the template to create resources with unique ids.
We can then choose a region close to us. At the time of writing we can choose between West US and West Europe, but I imagine this will grow once the Percept starts getting towards GA. I was actually surprised at the choice of regions here with no East US, North Europe, and no APAC region at all.
Moving on, the last item we need to select is the “LUIS prediction pricing tier, which we can leave at “Standard”, as the free tier doesn’t support speech requests sadly.
Hitting the “Create” button, then begins the process of deploying the speech theme resources.
We’re then prompted that this process can take between 2 and 4 minutes to complete…. This only took a matter of seconds for me nicely…
Azure Percept Audio – Hospitality Sample Template Demo
Once the template has completed deploying we’re then shown a demo Hospitality environment.
We should also now have 3 blue LEDs showing on the Percept;
I found these LEDs to be super bright, such that I couldn’t stare directly at them without then being able to see three (or is it 5? Ha) dots like Picard in that episode of Next Gen. They light my whole office up at night practically!
The Percept Audio LEDs will indicate different statuses depending upon their colour and flash pattern;
Ear SoM Status
|L02||1x white, static on||Power on|
|L02||1x white, 0.5 Hz flashing||Authentication in progress|
|L01 & L02 & L03||3x blue, static on||Waiting for keyword|
|L01 & L02 & L03||LED array flashing, 20fps||Listening or speaking|
|L01 & L02 & L03||LED array racing, 20fps||Thinking|
|L01 & L02 & L03||3x red, static on||Mute|
The LEDs are labelled as shown in the following picture, with L01 on the left of the SoM, L02 in the middle and L03 on the far right;
Returning to the Hospitality demo environment. The screen is split up into several sections.
At the top of the demo environment we have an toolbar containing;
- Create Custom Keyword
- Create Custom Command
- Get Started
- Learn More
Just below that we have the current keyword and command and links to change them should we wish…. We’ll actually be looking at all of that in another blog post, so keep your eyes peeled!
On the left we have an interaction area where we can enter commands for the Percept to action.
On the right we have a visual representation of the current environment, which reflects the actions our commands invoke.
Before we try executing any commands, the Percept uses the Speech Service to convert it’s command responses to spoken word.
For us to be able to hear that, we’ll need to connect some speakers to the device.
The Percept has a 3.5mm audio jack output for exactly that purpose… Hooking up some relatively low powered portable speakers to the line out jack will allow us to hear the responses to our commands.
We can now try executing some commands. The Custom Keyword or Wake Word for the Percept defaults to “Computer” (Where’s Scotty when you need him!), we can say that followed by one of a few commands which are applicable to this particular sample;
- Turn on/off the lights
- Turn on/off the TV.
- Turn on/off the AC.
- Open/close the blinds.
- Set temperature to X degrees. (X is the desired temperature, e.g. 75.)
I noticed, and perhaps due to my English accent, that it took a while for the Percept to recognise my pronunciation of “Computer”… I did try pronouncing it with an American Accent (and also asking it for the chemical formula for Plexiglass), but that didn’t seem to help.
Eventually it did work, and I quickly learnt how to say the word for a relatively repeatable wake up. I did notice that often, it would take quite a while to “Wake Up” the first time I issued the wake up word, and after that it would work quite quickly. I also noticed that, if I’d left it idle overnight that, in perfect IT Crowd style, I’d actually have to turn it off and back on to get it working again. When I get some time, I’ll raise some feedback with the team.
Once I’d mastered the wake word, all the other instructions worked pretty well…. You can see one of my failed attempts at the top here;
By instructing the Percept to turn on the TV, the simulation on the right would show the TV on, and so on through the commands.
The only command that didn’t work as intended was the “Set Temperature” command, which didn’t accept the actual temperature as a parameter to the command.
It turns out that this was because I was trying to set the temperature too low, you can only set it within a few degrees of the set temperature, otherwise it just doesn’t work.
The first thing I’m going to try with this is hook it all up to a Raspberry Pi and recreate the Hospitality experience in real life. I think this would make a really cool demo for future talks… (Also, Microsoft, if you need somebody to come and create you a demo room in Seattle, just shout eh! ;) Haha).
I did ask the team about perhaps using the Percept Audio to detect things other than speech. As I mentioned in my previous post, I have a client in the Ecology and Wildlife ecosystem, and I’d love to perhaps train the Percept to recognise the sounds of wildlife maybe?
Having spent time making Alexa Skills, Speech Studio is quite limited in comparison to the tools around Alexa, but it’s got everything we need at the moment to make reasonable speech based interaction apps.
I did find it frustrating that it would either not understand me, or be really slow to wake to the wake word… This makes demoing the unit a bit hit and miss and elicits the usual sniggers from attendees…. Ha.
Those points aside, I found the experience worked well, with the Sample Applications being a great example of some of the ideas Microsoft have in mind for this side of the Percept.
Brought to you by Dr. Ware, Microsoft Office 365 Silver Partner, Charleston SC.