Skip to content

README.md

Latest commit

 

History

History
executable file
·
45 lines (31 loc) · 2.74 KB

README.md

File metadata and controls

executable file
·
45 lines (31 loc) · 2.74 KB

Audio Detection Sample (gst-launch command line)

This sample demonstrates how to construct an audio event detection pipeline using the command-line utilitygst-launch-1.0.

How It Works

gst-launch-1.0 is a command-line utility included with the popular GStreamer media framework. It makes the construction and execution of media pipelines easy based on a simple and intuitive string format. Pipelines are represented as strings containing the names of GStreamer elements separated by exclamation marks !. Users can specify properties of an element using property=value pairs after an element name and before the next exclamation mark.

This sample builds a GStreamer pipeline using the following elements

  • filesrc or urisourcebin
  • decodebin3 for audio decoding
  • audioresample, audioconvert and audiomixer for converting and resizing audio input
  • gvaaudiodetect for audio event detection using ACLNet
  • gvametaconvert for converting ACLNet detection results into JSON for further processing and display
  • gvametapublish for printing detection results to stdout
  • fakesink for terminating the pipeline

Model

This sample uses the ACLNet model trained for audio event detection and made available through the Open Model Zoo. For more details see here.

  • aclnet_des_53_fp32.onnx is end-to-end convolutional neural network architecture for audio classification

NOTE: Before running this sample you'll need to download and prepare the model. Execute samples/download_omz_models.sh once to download and prepare models for all audio samples.

Model Proc

Along with the model network and weights, gvaudiodetect uses an additional model-proc json file that describes how to prepare the input for the model and interpret its output.

model-proc is a JSON file that describes the output layer name and label mapping (Cat, Dog, Baby Crying) for the output of the model. For each possible output of the model (specified using a zero based index) you can set a label and output specific threshold. Check here.

Running

./audio_event_detection.sh [INPUT_PATH]

Where [INPUT_PATH] can be:

  • local audio file ('./example.audio.wav')
  • network source (ex URL starting with http://)

By default, if no [INPUT_PATH] is specified, the sample uses a local file how_are_you_doing.wav (utilizing filesrc element).

Sample Output

The sample

  • prints full gst-launch-1.0 command to the console
  • starts the command and outputs audio detection results to the console