Records arbitrary length audio clips to send to Nvidia's ML inference server: Triton. Users can either read audio data from pulseaudio, which is configured in the client.conf file of this project, or pass a path to a directory containing audio clips to classify.

This container assumes that Triton and Pulseaudio are available on the node it is running on. Further, the microphone must be able to output sample at 16Khz.

• desc: Set to info, warning, or debug
• default: info

• desc: Number of classes to report
• flag: -c,--classes
• default: 5

• desc: Name of model in triton to perform audio inference against
• flag: -m,--model-name-classify
• default: ambient_sound_clf

• desc: Name of model in triton to generate audio embeddings
• flag: -e,--model-name-embeddings
• default: vggish

• desc: URL to access triton with
• flag: -u,--triton-url
• default: localhost:8000

• desc: url to access smarter-inference, default is empty string. If set, triton url will be overwritten within smarter-inference inference access point
• flag: --smarter-inference-url
• default: none

• desc: Seconds in between clip recordings
• flag: -p,--sound-poll-freq
• default: 10 seconds

• desc: Protocol to access triton with (HTTP or gRPC)
• flag: --protocol
• default: HTTP

• desc: How long each recorded clip should be
• flag: -r,--record-secs
• default: 10 seconds
• max: 40 seconds

• desc: Filepath for audio files to be used
• flag: --audio-file-dir
• default: /samples

• desc: Hostname for MQTT Broker
• flag: -b,--mqtt-broker-host
• default: fluent-bit

• desc: Hostname for MQTT Port
• flag: --mqtt-broker-port
• default: 1883

• desc: MQTT message topic string
• flag: --mqtt-topic
• default: /demo