I just finished the project Trigger Word Detection with model required in the description, but I still confused about the time distributed wrapper applied in the model. I have read the referred document about the introduction of time distributed wrapper and how it applied to dense layer which take the output of LSTM/GRU as input in Keras, and had applied it to build my word detection model.
How to Use the TimeDistributed Layer in Keras.
My question is how to map this implementation to a real world solution. Per my understanding a word detection model like Apple Siri, there would be a real-time microphone to detect what people is talking about and once a trigger word detect it response. While per the implementation of this model, it require a whole sequence of input be fed into the model, and a sequence of output to indicate when the trigger world detected. I want to know how this implementation with time distributed wrapper which require whole sequence applicable in a real-world product?
Hello Feihong,
Here’s a nice read that you can go through on basic implementation of a trigger word detection model. It basically shows an example of certain words picked and trained on Gated Recurrent Unit (GRU) using 3 types of audio recordings- Background, Positive words and Negative words.
In a real-time system, new samples from the microphone are continually added, and the oldest samples are discarded. The sample rate would be something like 8,000 samples per second, this is reasonably good for capturing voice audio.
The buffer that stores the samples needs to be large enough to store the longest phrase you are trying to detect.
At a slower rate (perhaps 10 times per second), the trigger detection process is run on the entire buffer.
Hi @Rashmi thanks for providing this and will go through the material!
Hey @TMosh is there any course relating to the system design during product management phase? so I need to take care of the metric you mentioned here? I’m planning to take MLops after this DLS, will it cover the issue?
@Feihong_YANG, will look forward to your response.
Some topics in MLops may contain some of what you’re looking for. I do not know because I have not attended the course myself.