The first step is to transcribe a set of human-human conversations.

# Maybe few ENV vars or config to use the right ASR/VAD

boot transcribe --audios-dir=<audios-dir> --conversations-json=<conversations-json>

--audios-dir is a directory with human-human stereo calls. One file per call. --conversations-json is where transcribed calls are kept as list of turns in the following structure:

{
  "audio-file-name": [<utterance>, ...], ...
}

<utterance>    := {"start": float, "end": float, "channel": str, "alternatives": [<alternative>], "text": str}
<alternative>  := {"confidence": float, "transcript": str}

boot train --conversations-json=<conversations-json> --output-model=<model.boot>

boot serve --model=<model.boot>             # Exposes an audio server
boot interact --model=<model.boot>          # Audio server working on CLI