ORAC: 'In-Universe' AI Voice & Terminal Chat

orac_test.mp4

ORAC-Voice is a fully autonomous, lore-accurate, voice-interactive AI, based on ORAC from the classic British Sci-Fi series Blake's 7.

I have always loved this show and have fond memories of watching it with my dad, when I was a kid. Jenna was always my favorite, with Avon taking second place. I always wanted my own ORAC, and this project is the first step in creating the brains, which hopefully, I will then put into a perspex chassis..

This project transforms a base model, Apple M4 Mac Mini into 'the' standalone, sentient super-computer. It listens continuously, processes speech locally using hardware acceleration, and responds strictly in the sardonic, impatient persona of ORAC, complete with an authentic sounding voice. (Subject to a little extra work and version of MacOS).

Features

• Authentic ORAC Voice: Using a "Custom" Apple Personal Voice. (Covered in Installation Steps 5 and 6, below.)
• Lore-Accurate Persona: 8K of Personality and Lore, covering important plot points from Season 1 of Blake's 7. ORAC is "in-universe" and will interpret the lore using second-person pronouns based on the character you choose. (Example, Q: 'How did we acquire the Liberator?' A: 'Following the failed mutiny aboard The London, You, Blake and Avon boarded the DSV-1....)
• Local Processing (No Cloud): Runs entirely locally on Apple Silicon for maximum privacy and low latency.
• Continuous Ambient Listening: Uses dynamic 'noise-floor' calibration to listen for voice commands without requiring a push-to-talk button.
• Barge-in Support: You can interrupt ORAC mid-sentence (via keyboard only by choice), and he will immediately halt his response and react.
• Operating Sounds: ORAC remains silent until you speak, which initiates the familiar 'key insertion' sound, a continuous 'hum/whirring' sound and finally when ORAC stops speak, the 'key removal' sound.
• Teletype UI: Custom ANSI terminal interface featuring live token tracking, memory monitoring, dynamic status lines, and scrolling history. (Independent control of teletype and voice speed.)
• Custom Phonetics Engine: A dedicated regex pipeline ensures ORAC pronounces the terminology (e.g., Servalan, Mutoids, DSV-2) with an appropriate RP accent.

Tech Stack & Hardware

This project is specifically designed to run on a dedicated Mac Mini M4 (16GB Unified Memory).

• LLM Backend: Ollama running the gemma2:9b-simpo model, providing excellent reasoning and adherence to system prompts. (Gemma3:4b seemed to hallucinate too much and other models are just too helpful!)
• Speech-to-Text (STT): MLX-Whisper (whisper-large-v3-turbo) optimized natively for Apple Silicon GPUs, paired with Python's SpeechRecognition library.
• Text-to-Speech (TTS): macOS native NSSpeechSynthesizer via PyObjC, utilizing an Apple Personal Voice clone. (Or SIRI voices if you don't wish to go through the Apple Personal Voice process.)
• Audio Processing: Native MacOS afplay for non-blocking UI sound effects. (Key and hum/whirring sounds.)

Project Structure

• orac_chat.py: The main asynchronous loop handling audio listening, the ANSI UI engine, subprocess sound loops, and Ollama inferencing.
• orac_data_core.py: The chronological history, lore parameters, and operational rules governing ORAC's knowledge base.
• orac_personality.py: Strict system prompts governing ORAC's arrogant tone, refusal to use filler words, and sardonic sign-offs.
• orac_phonetics.py: Regex dictionary that manipulates text strings before they hit the TTS engine to ensure proper sci-fi nomenclature and British intonations (e.g., Trap-Bath split: asking -> arsking).

Installation & Setup

Due to strict audio and accessibility sandboxing in recent macOS updates, this project relies on pyenv to manage a specific Python version, (3.12 Recommended), to ensure microphone and AppKit permissions function correctly.

1. Install System Dependencies: You will need Homebrew installed to grab portaudio(required for PyAudio/microphone access) & pyenv. (To use version 3.12 of Python in your venv).

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
# (Follow the instructions after install, to add Homebrew to your path.)

brew install portaudio pyenv

Instructions for setting up pyenv can be found here:

https://github.com/pyenv/pyenv?tab=readme-ov-file

(Follow the installation very carefully!)

2. Install Ollama and the model:

curl -fsSL https://ollama.com/install.sh | sh
ollama pull mannix/gemma2-9b-simpo:latest

3. Git Clone Project and enter Project Directory:

git clone https://github.com/CarolinaJones/orac-voice.git
cd orac-voice 

IMPORTANT: Stay in this directory for entire installation.

# Download python 3.12 using Pyenv and set it active locally:
pyenv install 3.12
pyenv local 3.12

# Create & Activate oral-venv
python3 -m venv orac-venv
source orac-venv/bin/activate

Ensure you are using pyenv to install and set Python 3.12. Do not use the default macOS system Python.

4. Install Required Python Libraries:

pip install mlx-whisper hf_transfer SpeechRecognition PyAudio ollama numpy PyObjC psutil

... next

curl -LsSf https://hf.co/cli/install.sh | bash
hf download mlx-community/whisper-large-v3-turbo --local-dir ./whisper/whisper-large-v3-turbo

You may wish to try this model as a smaller alternative:

hf download mlx-community/whisper-large-v3-turbo-q4 --local-dir ./whisper/whisper-turbo-q4

(Rememeber to udate the WHISPER_MODEL = in orac_chat.py if using this model.)

5. 'Hack' to allow Terminal to use Apple Personal Voice:

To authenicate Personal Voice in Terminal, paste this code in to your Terminal shell and press Return.

echo '#import <AVFoundation/AVFoundation.h>
int main(){ 
[AVSpeechSynthesizer requestPersonalVoiceAuthorizationWithCompletionHandler:^(AVSpeechSynthesisPersonalVoiceAuthorizationStatus status){ 
printf("Status: %ld\\n", (long)status); }]; 
[[NSRunLoop currentRunLoop] runUntilDate:[NSDate dateWithTimeIntervalSinceNow:2.0]]; 
return 0; }' > auth_check.m && gcc -framework AVFoundation -framework Foundation auth_check.m -o auth_check && ./auth_check

You'll get an authorization pop-up to agree to, and in MacOS settings, (under Personal Voice), you should now see that Terminal is authorized to use it.

6. Using Voicebox to train Apple Personal Voice:

Download latest Apple ARM release of Voicebox from:

https://voicebox.sh/#download

• Import the ORAC voice profile that I have included, in extras.
• In MacOS Settings, Accessibily, select Personal Voice to create a personal voice.
• You will be prompted to say ten phrase; Typically the first is, "I am creating a personal voice on my Mac."
• Generate the phrase in Voicebox with the ORAC voice profile, until you're happy with the result and play it back to Apple Personal Voice, create wizard.
• Repeat for subsequent phrases. (It might take a few goes!)
• Set your Mac's System Voice to your personal voice and make sure in the 'orac_chat.py', VOICE = "".

7. ...and now to configure some variables & test:

Ensuring you're in the directory, 'orac-voice', open 'orac_chat.py' in an editor, such as BBEdit and change these variables (ONLY) to suit.

VOICE = "" # Leave blank to use the "System Voice" - This allows for SIRI/Personal Voices. voice_pitch = 80.0 # Only works on SYNTH voices and not SIRI/Personal voices.

U1 = 0.06 (Teletype Speed) U2 = 0.071 (Teletype Uniformity)

TRANSCRIPT_DIR = /Users/Caroline/Desktop/ (A Directory called transcripts will be created.) TR = "ORAC_Transcript_CM" # Transcript Name Prefix (Date will be added).

YOUR_NAME = "Jenna" # USER Name and Identity

ORAC_NAME = "ORAC" # ORAC's Name

TERMINAL SETTINGS:

TERMINAL_PROFILE = "Homebrew" # Terminal Profile

TERMINAL_FONT = "AdwaitaMono Nerd Font Mono" # Font Name (This font is in extras)

TERMINAL_FONT_SIZE = 17 # Font Size

TERMINAL_COLS = 100 # Window Width

TERMINAL_ROWS = 35 # Window Height

..and then (From orac-voice Folder):

source orac-venv/bin/activate
python3 orac_voice.py

• Boot Sequence: The Terminal will resize, display a booting animation, and calibrate to your room's ambient noise floor. (Updates dynamically throughout conversation.)

• Interacting: Address ORAC naturally. The system is voice-activated but ignores background noise. Press ESC to interrupt. (Give him 9-10 seconds after start-up to load the model to RAM. Once loaded you should hear a response within 1-2 seconds. The model stays loaded in memory for 3 hours.)

• Keyboard Entry: You can manually type text into the bottom UI bar. (Typing and entering text while ORAC is talking, will "barge-in".) While ORAC is not actively talking, you can use arrow keys or mouse scroll to review your session. Use FN key + CMD+ C & V as usual, for copy and paste.

• Purge Memory: Say or type "Clear memory" and your available tokens will refresh. Keep in mind, the relative low context window of this model - I have set a "sliding" option, when around 85% of your tokens are used.. By default hisrory is set to twelve, this will remove two from the top to allow continuous conversation.

• Header Status: The Header is colour-coded in a traffice light scheme, with status, to let you know how many tokens you've used.. Below is a TKNS: xxxx/8192 and RAM: xx% (Tokens available out of 8192 and RAM useage of your Mac.)

• Shut Down: Say "shut down" or press Ctrl+C. ORAC will prompt you to save a transcript of the conversation before powering off, or to cancel and return to your conversation.

Acknowledgements

• The estate of Terry Nation and the BBC for the enduring legacy of Blake's 7.
• Peter Tuddenham, for providing the unforgettable original voice of ORAC.

I hope you enjoy! Caroline xo