Listen for a live recording and transcribe the file. Send a notification when YOUR_NAME is called

transcribe.js

const pathParser = require("path");
const { MODEL_PATH, HOME, YOUR_NAME, FILE_GLOB } = process.env;
console.log(MODEL_PATH, HOME, YOUR_NAME, FILE_GLOB);
const chokidar = require('chokidar');
const vosk = require('vosk');
const wav = require("wav");
const { Readable } = require("stream");
const { notify } = require('node-notifier');
const TailingReadableStream = require('tailing-stream');
const recordingsDir = `${HOME}/recordings/`
const model = new vosk.Model(MODEL_PATH)
const listenPath = `${recordingsDir}${FILE_GLOB}`
console.log('listenPath', listenPath)
const watcher = chokidar.watch(listenPath, { persistent: true });
// const Speaker = require('speaker');
const audioStreams = {}
let startup = true;

watcher
  .on('add', async (path) => {
    if (startup) console.log('skipping due to startup ')
    if (startup) return
    const file = pathParser.parse(path)
    // expects files to be named with J prefix eg. J2-blah-blah.wav
    const [jHost] = file.name.split('-')
    console.log(`New recording ${jHost} ${file.base}`)
    const wfReader = new wav.Reader();
    const wfReadable = new Readable().wrap(wfReader);
    console.log(`about to setup ${jHost} ${file.base}`)
    wfReader.on('format', async (format) => {
      const { audioFormat, sampleRate, channels } = format
      const rec = new vosk.Recognizer({ model: model, sampleRate: sampleRate });
      // const speaker = new Speaker(format)
      if (audioFormat != 1 || channels != 1) {
        console.error("Audio file must be WAV format mono PCM.");
        return
      }
      for await (const data of wfReadable) {
        const end_of_speech = await rec.acceptWaveformAsync(data)
        if (end_of_speech) {
          const { partial } = rec.partialResult()
          console.log(partial);
          rec.reset()
          if (partial && partial.includes(YOUR_NAME)) {
            notify({
              message: `${jHost} said your name`,
              sound: 'Glass'
            })
            // hear the output
            // wfReader.pipe(speaker)
          }
        }
      }
      rec.free();
    });
    // nodejs fs module will stop when it gets to EOF, so we use TailingReadableStream to follow the tail of the file
    const stream = TailingReadableStream.createReadStream(path, { timeout: 0, highWaterMark: 4096 });
    stream.pipe(wfReader)
    audioStreams[path] = stream
  })
  .on('change', async path => {
    // file changed, likey finished with the file
    if (audioStreams[path]) audioStreams[path].destroy()
  })

setTimeout(() => {
  startup = false
}, 5000)

This will listen for changes in your recordings directory and transcribe wav files as they are being written to. Listens for YOUR_NAME and pops up with an alert of which Job said your name.

Install dependencies:

npm install chokidar vosk wav node-notifier tailing-stream stream speaker

You'll also need to save the model to the model directory in the same directory as trasnscribe.js. Find the models here

Record audio using timemachine. I have a custom build using PCM 16 bit audio as the output which would be required for this to work properly.

timemachine -t 1 -T 15 -p $HOME/recordings/J1/J1- -o 7132 -n J1_record -f wav -a J1:receive_1
timemachine -t 1 -T 15 -p $HOME/recordings/J2/J2- -o 7132 -n J2_record -f wav -a J2:receive_1

Files require a particular format to identify which job is saying your name.

$HOME/recordings/J1/J1-blah-blah.wav

Run like this

MODEL_PATH="model" YOUR_NAME="nathan" FILE_GLOB="**/*.wav" node transcribe.js