Windows Voice Commands Batch Script

Clunky but private and configurable system for voice commands on Windows using batch scripts. It is not very fast but I find it useful enough for running commands while not sitting at PC or to not interrupt current tasks.

To overcome bad piping support in Windows, this script makes voice capture and voice recognition dependencies communicate by writing files.

In essence: this routes first spoken word to a batch script of the same name and provides the rest of the sentence as a variable to do with as you please.

Dependencies

ffmpeg for capturing input. May be replaced by altering line 8 of wvc_run.bat. The replacement must output wvc.m4a file in the same directory.
OpenAI Whisper for voice recognition run locally. May be replaced by altering line 9 of wvc_run.bat. The replacement must output the text without punctuation symbols into wvc.txt file in the same directory. One easy way to install Whisper is with Python 3 and pip pip install -U openai-whisper.
It is recommended to use some advanced solution for keyboard shortcuts such as Powertoys to create a keyboard shortcut that will start the wvc_run.bat.

Configuration

Basic parameters can be configured inside wvc_config.bat.

You will definitely need to replace the wvc_recording_device with your audio input device, find its name in the "Manage Audio Devices" settings. For most cases replacing the part inside brackets with the device name will be enough.

The small Whisper model is small (~500MB) and good enough for English language use. If you really need more precision and support for different languages replace the wvc_whisper_model setting with the model name and wvc_whisper_language with a language supported by the chosen model.

On Windows you may run into SSL error when fetching the model for the first time. You can spend hours fixing that or just manually download the model and then set wvc_options to --model_dir C:\path\to\model where the path needs to point to directory where model is downloaded and not to the file itself. "Small" model can be downloaded directly from direct download

Other configuration options should be self-explanatory.

Built-in Commands

You are meant to create your own commands, but there are some example built-in commands, currently:

"Google {topic}" will open Chrome and google for a topic.
"Run {program name}" edit the wvc_commands/run.bat to add running logic for programs you need.

Custom Commands

Built-in commands are really just examples, you are encouraged to create your own by adding a batch script into wvc_commands folder with file name matching the command, i.e. find.bat. Only one word is supported for file matching, but you can add logic into script to match more words, see run.bat for example.

%wvc_subject% holds all the words after the command, it is up to you to parse and use it.

Usage

Run the wvc_run.bat script either manually or with keyboard shortcut. Mind that you must run it as a user to have all the paths in environment available to that user.

Now comes the clunky part. You must wait about half a second for ffmpeg to display that it is receiving input from microphone, and then you have 5 seconds (depending on configuration in wvc_config.bat) to speak your command, after which it is sent to voice recognition. If you are running a model for the first time, Whisper will have to download it first. Otherwise, it will take couple more seconds (more or less, depending on your hardware) for command to be processed into text and run.