Custom Sound Detection

[Jace Ford]

I apologize if this is in the wrong forum. I have a project in mind that I need any thoughts, ideas, suggections and collaboration.

First, the general idea: I want to build a device that can be setup outdoors that listens at all times and when a particular sound is heard it recognizes it and sends a notification that the sound was heard. And ideally it would have multiple microphones that could be used to detect sound direction as well. Think amazon alexa but with a custom sound detection. I know their are devices that do this for gun shots that law enforcement uses to help identify where the shot came from and potentially what type of weapon created the sound. For example the device could be placed in the woods and listen for a particular bird chirping or a turkey gobble or a coyote howl or owl hoot. Could use that info to surveille for a particular species in an area and what time of day they may be active. Could also be used to listen for a vehicle in an area that is off limits to vehicles or a gun shot in a no hunting zone....

Progress thus far: Initial searches continue to pull up new custom sound features that Amazon Alexa's have now. Or speech recognition for that latest phones This makes finding the right keyword search somehwhat difficult

Equipment: I have found several speech recognition devices and noise detection devicesNot sure if those devices could be used or not for listening for a custom sound. Really a custom app for an Alexa might would be ideal if paired with a power source and a weather proof enclosure that would still allow for sound to reach the device. An Alexa or Google Home.... already has the microphones, and network interface components and custom software can be written for them.

Software: I realize this is probably the most difficult part. However I also realize this doesn't have to be 100% accurate and could be a work in progress unlike something that may be used in public safety. I'm a little uncertain how to upload all the variations of a particular sound to the device for recognition.

Thats the project in a nutshell, anyone wanting to help is certainly welcomed. Thanks.

 

[Bluejets]

Almost impossible to detect a particular sound.

 

[Harald Kapp]

I say it depends on the sound.
Here's an example of a gift box that open only if the right phrase is spoken. It uses a Raspberry Pi zero for voice recognition. The same software may, methinks, the trained to detect any other specific sound. You won't need the woodworking part of that project. Instead you will have to add the notification part (WiFi, Bluetooth, SMS, Lora, ...?).

 

[kellys_eye]

As Harald says, voice (sound) recognition is pretty well advanced and achieving what you require is certainly possible but not without considerable effort and research/design.

Dual (possibly more) microphones (for direction sensing), audio filtering (to eliminate as many non-required sound sources as possible) and software decoding of the signal are but three of the processes you'd have to cope with to get the end result.

What you would end up with is a (potentially) commercially profitable solution but not without considerable input and without knowing what your own contribution to that would be I doubt if there's anyone on this forum that would consider putting (potentially) 90%+ of the required effort into it - and for what return?

 

[Nanren888]

There is lots of work in this sort of area. Distributed sensor networks are often another way to approach the localisation problem: Trade complicated direction-capable sensors for known locations of multiple sensors.
I have seen several vaguely similar projects in New Zealand. A coule of the links towards local work in the area.
LOcalisation of bird calls in NZ bush is a link to a page, that has a few of Yusuke's references on the topic. Can't remember who ran that work before he came.
Temporal and environmental influences on the vocal behaviour of a nocturnal bird, This guy Paul Teal is broadly extraordinary and also has worked in this area.
.
Algorithms for detection or direction of arrival are generally not the issue.
Aside: Some of Yusuke's other papers might suggest that for a known environment, you might even get a range estimate without a long baseline of mics.
As suggested above, some practical things are likely more the problem.
Algorithms tend to fit into a couple of categories. Either blind and require a large amount of data to train and training is slow, or take a lot of feature selection and design.
Most algorithms are computationally intensive, hence lots of network traffic, or local/edge compute capability.
I'm surprised that you have network connectivity: often a problem. Not clear what you want to do with this capability, as that often determines the connectivity you need.
More compute capability generally means more power supply.
Blind source separation may help if the animal life is dense, overlapping. But requires more computation.
.
Simple at the edge detection is sometimes used in vision solutions to get an idea of when to record, or when to forward data, to reduce storage requirements over recording everything. That is to layer recognition and storage or processing. It might be worth considering.
.
But, wind and ambient noise is often an issue: it's windy here. If you don't know this area, Google "dead cat".
Weather proofing and "dead cats" don't always go together.
Power supply? Maybe that's easy if you also have network? Platforms capable of recognition are often not the lowest of power.
Algorithm development: can be treated by some as easy enough, say N-Nets, deep-learning, but needs significant data and compute resources, or if designed, eg SVMs & such, significant design effort. If not your area Google "machine learning" or scikit-learn.
Hardware design also likely costly and time-consuming. As mentioned above, audio filtering, just plain audio quality, compute capability, power supply, networking. Eventually, may need to have a mechanism to locate the deployed units. Might want to prove capability before embarking on prototyping deployable hardware.
.
As the University of Auckland and Wellington links show, it's often an active and feel-good area in many places, so there are options to get a local university student project: sometimes these can be free resource and even sometimes provide steps towards solutions. But, this means you can usually track down similar research, if you prefer.
.
Sorry, it's mostly cautionary notes.
Hope something helps.
Enjoy