Skip to content

Latest commit

 

History

History
110 lines (87 loc) · 5.57 KB

README.md

File metadata and controls

110 lines (87 loc) · 5.57 KB
Raw

MMM-WhisperGPT

This is a module for the MagicMirror².

IMAGE ALT TEXT HERE

How it works 👉 https://nikro.me/articles/professional/crafting-our-ai-assistant/

Goal of the module is to create a custom interactive widget that uses Open AI tools:

  • Whisper - self-hosted model for voice-to-text transcription.
  • LangChain - intended to be used with ChatGPT API, to process the requests.
  • Picovoice -> Porcupine - is used for offline (self-hosted) word trigger (accent on the privacy).
  • also... mimic3 :)

Idea is the following:

  1. Wake word (Porcupine).
  2. ...record query (show a sexy animation, will be done later)
  3. ...pass to self-hosted Whisper
  4. ...transcribe voice-to-text
  5. Show the question as transcribed rendered-text (in the module render)
  6. ...pass through LangChain to ChatGPT
  7. ...pass the textual reply back to the module and render on-screen
  8. ...use TTS (mimic3) - self-hosted on the network, to throw back a wav file to play.

Using the module

To use this module, add the following configuration block to the modules array in the config/config.js file:

var config = {
    modules: [
        {
            module: 'MMM-WhisperGPT',
            config: {
                // See below for configurable options
                picovoiceKey: 'xxx',
                picovoiceWord: 'JARVIS',
                picovoiceSilenceTime: 3,
                picovoiceSilenceThreshold: 600,
                audioDeviceIndex: 3,
                openAiKey: 'xxx',
                openAiSystemMsg: 'xxx',
                whisperUrl: '192.168.1.5:9000/asr',
                whisperMethod: 'openai-whisper',
                mimic3Url: '192.168.1.6:59125'
            }
        }
    ]
}

Configuration options

Option Required? Description
picovoiceKey Required Picovoice access key - you have to register to obtain it - this is used for trigger word.
picovoiceWord Optional Picovoice trigger word, i.e. BUMBLEBEE, JARVIS, etc. Defaults to JARVIS.
picovoiceSilenceTime Optional Silence period - defaults to 3 (3 seconds).
picovoiceSilenceThreshold Optional This is usually background noise * THIS NUMBER. Default value is 1.1 (aka 10%).
audioDeviceIndex Optional Audio device - i.e. 3 - those will be printed out when you're using debug mode. Defaults to 0.
whisperUrl Required URL (or IP?) to self-hosted instance of the Whisper.
whisperMethod Optional Whisper method: openai-whisper or faster-whisper. Defaults to: faster-whisper.
whisperLanguage Optional Defaults to: en.
openAiKey Required API Key of OpenAI.
openAiSystemMsg Optional System msg - how the AI should behave.
mimic3Url Required Mimic3 URL (server), with protocol, port, without /api/tts
mimic3Voice Optional Mimic3 Voice - default: en_US/cmu-arctic_low%23gka
debug Optional If you want to debug, default is: false.

What is Picovoice / Porcupine

Picovoice / Porcupine is used for the "Trigger" word. It's a self-hosted small AI / Neural Network (NN). Picovoice offers a range of services, including a license for this offline AI. It only sends usage statistics, not the actual audio conversations.

What is Whisper

Whisper is an open-source product from OpenAI. It's a Large Language Model (LLM) AI that handles speech-to-text (transcription). In my personal case, I have it self-hosted on my local network.

I used this: https://github.com/ahmetoner/whisper-asr-webservice

What is ChatGPT

ChatGPT is another product from OpenAI. It's a Large Language Model (LLM) AI. You will need to register and get an API Key to use it.

What is LangChain

LangChain is a library built around LLMs that allows for extra functionality, such as long-term memory.

What is Mimic3 (Mycroft)

Mycroft's Mimic3 is a Text-to-Speech (TTS) system based on a Large Language Model (LLM). It offers realistic TTS that can run on somewhat resource-restricted systems. I initially tried to set it up on my OrangePi, but instead, I installed it on the same machine with Whisper and use it via the network.

I used this docker-compose.yml 😉

version: '3.7'

services:
  mimic3:
    image: mycroftai/mimic3
    ports:
      - 59125:59125
    volumes:
      - .:/home/mimic3/.local/share/mycroft/mimic3
    stdin_open: true
    tty: true

Troubleshooting

  1. If your audio doesn't work - check if you're using alsa or pulseaudio. You might need to install mpg123. You can install it using the command sudo apt-get install mpg123.
  2. You might also need to install lame for audio encoding. You can install it using the command sudo apt-get install lame.