You will need to install ffmpeg for the transcript-worker to function. This handles getting input across many different channels as well as deal with setting sound thresholds and such. To get it, for macOS you should use homebrew and for Linux use your distro's package manager. Alternatively for those platforms, and for Windows generally, you can directly download the necessary binaries and put them somewhere on your path.
Then, you need to get the STT Credentials and then do:
npm install
cp cog-sample.json cog.jsonnode transcribe.jsYou can go to http://localhost:4545 to view a web UI that allows you to on-the-fly configure channels and view their current status.
To run this worker, you will need a cog.json file to handle its configuration. Discussed below are the various elements that go into it.
Before continuing, you should determine the device that you want to use for the worker.
To find out device names:
On Windows, use: ffmpeg -list_devices true -f dshow -i dummy.
On Mac, use: ffmpeg -list_devices true -f avfoundation -i dummy.
On Linux, use: arecord -L.
For Mac and Linux, you can probably just use default, and choose the device through the system dialog. On Windows,
you may need to take the name of the device from the command above and set the device key for a channel to
point to that device, for example:
{
// other settings
"transcribe": {
// other transcribe settings
"channels": [
{
"device": "Microphone (2- USB PnP Audio Device)"
// other channel settings
}
]
}
}If you wish to utilize RabbitMQ for the worker, you must specify the following:
{
"rabbit": true
}If this is not specified, the transcript worker, will just write each incoming transcription to the console, and nothing else. Refer to documentation for @cisl/io for further configuration details. See below for the RabbitMQ topics and payloads for this worker.
You must input the credentials of Watson Speech-to-Text service,
utilizing the ibm-credentials.env file for the service. See
node-sdk#getting-credentials
for details on how to get the file and details about it.
Configuration for running the transcript-worker happens under the transcribe key in the
cog.json file. Many of the values can be omitted, and will be instantiated with the defaults
described below.
Next, you will need to configure the default details of the service:
{
"transcribe": {
"default_driver": "ffmpeg",
"default_device": "default",
"default_language": "en-US",
"default_model": "broad",
}
}This provides default settings for each channel you specify (see below). The device should be set using the instructions above. The language should be the IETF format with ISO 3166-1 country code. The model should either be "broad" or "narrow". You can go to https://console.bluemix.net/docs/services/speech-to-text/models.html#models to see the full list of support languages and models.
The driver can be either ffmpeg (to allow piping from ffmpeg onto that channel) or
fake (only allow typing in inputs from web console).
Note: Leaving these out of the configuration will use the values shown above.
Channel configuration is handled by a list of objects as shown below:
{
"transcribe": {
"channels": [
{
"device": "device name",
"language": "en-US",
"model": "broad"
},
{
"_comment": "use default settings",
"driver": "fake"
}
]
}
}The first channel has all values explicitly defined with the second channel using the defaults (see above). You can specify as many or as few of the properties for a channel as you want.
Note: If you don't include the "channels" key, then the system will default to one channel with the default settings.
{
"buffer_size": 512000,
"speaker_id_duration": 30000,
"max_alternatives": 3,
"publish_interim": false
}The buffer_size specifies how big of a stored buffer each channel will have for the incoming binary
sound buffer from ffmpeg, allowing one to fetch this for later analysis as requested.
The speaker_id_duration is how long should a channel save a speaker id once specified. If set to
false, then speaker ids will never be removed.
The max_alternatives sets how many possible alternative transcriptions is returned by the Watson service.
The publish_interim setting when true will output all interim results as they are received from the Watson
STT service. Setting it to false will output only the first interim result, and no subsequent values.
Note: Leaving these out will use the above default values.
Assuming you've configured the worker to use RabbitMQ as specified above, the worker will publish its results to RabbitMQ as well as allow you to configure it.
The transcript-worker as it receives results publishes to two topics in RabbitMQ:
- transcript.result.interim
- transcript.result.final
The first topic is broadcasted by the worker as the speech to text service attempts to figure out what was said. This is useful for UI elements to show the user that the system is listening to them. The second topic is the final transcription returned by the service and should be taken as what the user said to the system.
Both of these topics have the following message structure:
{
"worker_id": "transcript-worker",
"message_id": "5a5b5940-1f33-11e9-b7f9-23b47ef98a1b",
"timestamp": "2019-01-23T17:21:43.124Z",
"channel_idx": 0,
"transcript": "yeah range",
"total_time": 0.97,
"result": {
"alternatives": [
{
"timestamps": [
[
"yeah",
2.87,
3.21
],
[
"range",
3.43,
3.84
]
],
"confidence": 0.648,
"transcript": "yeah range"
}
],
"final": true
}
}The transcript worker listens for a range of commands to change how it's operating or to request some sort of data on the next input. These commands are:
- transcript.command.tag_channel:
{"channel_idx": number, "speaker": string} - transcript.command.switch_language:
{"channel_idx": number|null, "language": string} - transcript.command.pause:
{"channel_idx": number|null} - transcript.command.unpause:
{"channel_idx": number|null} - transcript.command.extract_pitchtone
{"channel_idx": number} - transcript.command.start_publish
{} - transcript.command.stop_publish
{}
Additionally, the following legacy topics are supported:
- switchModel.transcript.command:
"language_model_name" - switchAcousticModel.transcript.command:
"acoustic_model_name" - controlAudioCapture.transcript.command:
{"channelIndex": number|null, "command": "pause"|"unpause"} - rpc-transcript-|config.id|-tagChannel (RPC Queue):
{"channelIndex": number, "speaker": string}
NOTE: You can use channel_idx and channelIndex interchangeably,
however, it is recommended to move to channel_idx as channelIndex will be
removed at some point.
This application does not directly interact with the Learning Assistant,
but rather exposes a binary buffer along the transcript.pitchtone.audio topic (after extraction
was requested for a channel using transcript.command.extract_pitchtone). After requesting an extraction, the
next spoken word/phrase that is transcribed will be sent, regardless of the usage of a wake-up word or otherwise.
Below is an example of code that can be used to interface with the transcript-worker and Learning Assistant:
const io = require('@cisl/io');
const wav = require('wav');
const fs = require('fs');
const fetch = require('node-fetch');
const FormData = require('form-data');
io.rabbit.onTopic('transcript.pitchtone.audio', (response) => {
const file_name = `output-${io.generateUuid()}.wav`;
var save_wav = new wav.FileWriter(file_name, {
channels: 1,
sampleRate: 16000,
bitDepth: 16
});
save_wav.on('finish', () => {
const form = new FormData();
form.append('format', 'json');
form.append('sample_rate', '16000');
form.append('file', fs.createReadStream(file_name));
form.append('word', '校内');
fetch('https://internal.cisl.rpi.edu/la/upload', {
method: 'POST',
body: form
})
.then(res => res.json())
.then(json => {
fs.unlinkSync(file_name);
console.log(json)
});
});
save_wav.write(response.content);
save_wav.end();
});
io.rabbit.publishTopic('transcript.command.extract_pitchtone', {channel_idx: 0});where the output to the console will be:
{
"filename": "<file_name>"
}which corresponds to https://internal.cisl.rpi.edu/la/renders/<file_name>, for example:
{
"filename": "1570115937_855945.webm"
}corresponds to https://internal.cisl.rpi.edu/la/renders/1570115937_855945.webm.