This project lets you transcribe your audio input to Google Docs using Google's Cloud Speech API. You can edit the document in realtime (with a lag of a few seconds). If you don't need the ability to edit transcripts in realtime, you may prefer to use one of the many other, more polished transcription services.

The Cloud Speech API is not free, so you will need to provide your own credentials to use this project.

As with any node project, npm install to install dependencies.

You will also need to install sox and have it available on your $PATH. On MacOS you can install it with Homebrew as brew install sox.

Create a Google Cloud project, enable the Speech-To-Text API for that project, create a service account, and aquire private key for that service account. You can do that all in one step by clicking the "Set up a project" button behind this link. Then copy the JSON with the private key for the service account to a file named speech-service-account-key.json in this directory.

Enable the Google Docs API for your project. You can do this from the APIs & Services page on the Cloud console. Make sure you're in the right project at the top, then click "Enable APIs and Services", search for the Google Docs API, and click Enable.

Set up the app to allow users to grant it permission to edit documents on their behalf. From the APIs & Services page, click "OAuth Consent Screen" in the left menu, click "External", and press "Create". Provide a name and email to be displayed to users and a developer contact, then click Save and Continue. On the Scopes page, select the ".../auth/documents" scope ("View and manage your Google Docs documents"). Ignore the warning about verification for now. Click Save and Continue until you get to the Summary page. Don't worry about registration for now, just click Back to Dashboard.

Finally, acquire the OAuth secret for your application. From the APIs & Services, select "Credentials", click Create Credentials at the top, select "OAuth Client ID", set the type to "Desktop app", give it a name, and click "create". It will tell you the ID and secret, but it's easier to click through and the click the download icon to the right of the newly created credential. Copy that JSON to a file named gdocs-client-oauth-secret.json in this directory.

The file replacements.js defines a function which transforms snippets as they come in, which is useful if the transcription makes consistent mistakes. The current set of replacements is optimized for TC39's needs, so you will almost certainly want to replace it.

Changes to that file are loaded automatically; you don't need to restart the transcription.

node run.js

This will ask for the ID of a Google Doc to transcribe to. This is the 1fOSJp9hZLoR2BhiE0da87bLtoBBgcvGdi063IU1wqW8 part of doc's URL, as in https://docs.google.com/document/d/1fOSJp9hZLoR2BhiE0da87bLtoBBgcvGdi063IU1wqW8/edit. Make sure it's a document you have permission to edit.

If this is the first time running it, you'll then be asked to authorize an account. (The resulting credential is stored in GENERATED_TOKEN.json, so you should only need to do this once.) Go the URL it provides and grant permissions. You'll need to click through the big "This app isn't verified" warning (under "advanced"), unless you've verified your application. Ultimately you should get a string you will need to copy back to the terminal.

That's it! It will then run forever, transcribing your microphone to the doc. Kill it with control-C when you're done.

Keep in mind that this costs money. Specifically, since this is configured to use the fancy "video" model (which seemed to work best in my experience), it will cost $0.009 every 15 seconds, or about $2.16 an hour.

This records audio in, not audio out. So it won't directly help with, e.g., transcribing a stream you're viewing on your computer. You'll need to use a tool like JACK or BlackHole (MacOS only) to route your audio out to your audio in.

The point of this is to have realtime transcriptions. My use case for that is so someone can be fixing up errors made by the STT service while they're in the meeting (or whatever) being transcribed, instead of needing to a whole additional chunk of time fixing it after the meeting (which may not be practical if a recording is not available).

Unfortunately there's an inherent delay from the STT service, which means the transcript is necessarily going to be behind, so you have to constantly keep the last ~ten seconds in short-term memory. That's doable, but you may find it helpful to use a tool like Rogue Amoeba's Audio Hijack to delay the audio from the meeting by a few seconds, to match the transcription. (That particular tool is Mac only. The feature you need is available in the unlimited free trial.) This is probably annoying if you're going to be speaking frequently, so I recommend only making use of this during portions of the meeting when you're not talking on the call.