This is a python app demonstrating On-Device Streaming ASR using QuartzNet model executed with Intel(r) OpenVINO

The app will continuously stream audio coming through microphone, break it into chunks and some transformation before sending it to the neural network (QuartzNet-15x5) to get character probabilites. Subsequently, perform CTC decoding and optionally apply KenLM to improve WER.

Intel(r) Distribution of OpenVINO 2022.1

The list of models supported by the demo is in models.lst file. This file can be used as a parameter for Model Downloader and Converter to download and, if necessary, convert models to OpenVINO IR format (*.xml + *.bin).

An example of using the Model Downloader:

omz_downloader --list models.lst

An example of using the Model Converter:

omz_converter --list models.lst

• quartznet-15x5-en

NOTE: Refer to the tables Intel's Pre-Trained Models Device Support and Public Pre-Trained Models Device Support for the details on models inference support at different devices.

Run the application with -h option to see help message.

usage: streaming_asr_openvino.py [-h] -m MODEL [-d DEVICE]

optional arguments:
  -h, --help            Show this help message and exit.
  -m MODEL, --model MODEL
                        Required. Path to an .xml file with a trained model.
  -d DEVICE, --device DEVICE
                        Optional. Specify the target device to infer on, for
                        example: CPU, GPU, HDDL, MYRIAD or HETERO. The
                        demo will look for a suitable OpenVINO Runtime plugin for this
                        device. Default value is CPU.

The typical command line is:

python3 streaming_asr_openvino.py -m quartznet-15x5-en.xml

The application prints the decoded text for the audio coming through the mic in real-time until the program is terminated.

• Open Model Zoo Demos
• Model Optimizer
• Model Downloader