BD-1 Voice Assistant Project

Welcome to the BD-1 Voice Assistant Project! 🎧🤖

This repository aims to build a custom Text-to-Speech (TTS) system that mimics the vocal style of BD-1, the adorable droid from Star Wars Jedi: Fallen Order. The TTS reacts to user input using pre-recorded BD-1 sounds that are matched to emotion and sentence structure.

Work in Progress: The TTS core is functional. Emotion detection and assistant logic are under improvement.

---

Project Goal

To build a TTS system in BD-1's style, capable of:

• Responding to user prompts using a combination of BD-1 sounds.
• Matching emotion (e.g. happy, sad, surprised) to the content of the phrase.
• Mimicking language structure while remaining robotic and cute.

---

Before starting

pydub seems to have dependencies that do not work with Python 3.13. I provide a fix_pydub.py to run before doing anything.

---

Current To-Do List

Here's what's left:

• Improve emotion detection from text input. (see assign_emotion() function)
• Test different styles and expressions with more BD-1 phrases.
• Find a way to make BD-1's sentence shorter and keeping a "grammatical" sens

---

📂 Folder Structure

BD-1-Conversationnal-AI/
├── fix_pydub.py               # First script to run
├── text_to_speech_v0.py       # Core logic for BD-1 voice synthesis
├── text_to_speech_vX.py       # (latest)** → Production-ready version, callable from other scripts without flooding logs.
├── test.py                    # Main script to test BD-1 voice playback
├── get_sound_chunked2.py      # Another version of get_sound_chunk that look for more than one audio file per word
├── sounds/                    # Find more info in the Readme.md in the sounds folder.
│   ├── consonnes/             # Raw consonant sounds (neutral)
│   ├── emotions/              # Same sounds, sorted by emotion (happy, sad, etc.)
│   └── compositions/          # Multi-letter chunks classified by sound families and emotion
│       ├── 2_caracteres/      # Composition of 2 sounds
│       ├── 3_caracteres/      # Composition of 3 sounds
│       └── 4_caracteres/      # Composition of 4 sounds
└── temp_tts.wav               # Temporary output file (should be deleted everytime)

Functional Breakdown (text_to_speech_vX.py)

tts_bd1(message: str)

Main entry point to play a complete BD-1 response.

• Splits full message into emotional phrases
• Calls generate_tts_audio() per phrase
• Assembles and plays final audio

process_message_by_phrases(message: str)

Splits the text into short sentences using ., !, ? and detects emotion per phrase.

• Returns a list of (phrase, emotion).
• assign_emotion(phrase: str)

Rules-based emotion classifier:

Keyword	Emotion
?	Question
!	Surprise
Keywords like non	negatif
Keywords like oui	positif
Keywords like triste	triste (sad)
Else	neutre (neutral)

decompose_message(message: str)

Extracts consonants by syllable:

• Removes vowels and accents
• Only keeps consonants followed by vowels
• Groups leading consonants (e.g. dr, gn, mp)
• Discards words of length ≤ 4 unless exception. If no word over 4 letters, accepts 4 letters words.
• Returns a cleaned list of consonants.

map_letters_to_sound_groups(text: list[str])

Converts consonants to families:

• B = Beep
• S = Sifflement (whistle)
• P = Piano

get_sound_chunked(consonnes, emotion)

Core of audio matching logic.

• Tries longest chunks (4 → 3 → 2 → 1)
• Uses folders like:
• sounds/compositions/3_caracteres/Beep Piano Sifflement/positif Fallback:
  • Same chunk in neutre
  • Then letter-level fallback with ### get_sound()
• Ensures no .wav is used twice
• Returns list of (path, emotion, chunk_length, original_chars)
• CURRENT VERSION each word is limited to one chunk or character to avoid long message

get_sound(consonne, emotion)

Fallback sound lookup:

• Tries in sounds/emotions/{emotion}/
• If not found, tries sounds/consonnes/ (Neutral sound folder)

generate_tts_audio(message: str, options)

Assembles final audio from selected chunks:

• Uses AudioSegment.from_wav
• Displays full table of sound matches
• Returns in-memory WAV file

What Happens Step-by-Step

Suppose the message is:

"Bonjour à toi ! Comment vas-tu ?"

1. tts_bd1() calls process_message_by_phrases() Breaks into:

• Phrase 1: "Bonjour à toi !" → surprise
• Phrase 2: "Comment vas-tu ?" → question

2. For each phrase: generate_tts_audio() is called

3. decompose_message() extracts consonants Keeps only relevant syllables :

• ['b', 'n', 'j', ' '] ignore other words (too small)
• ['c', 'm', 'm', ' '] ignore other words (too small)

4. map_letters_to_sound_groups() is called Convert letters to sound group to find chunks :

• ['b', 'n', 'j', ' '] => ['B', 'P', 'S', ' ']
• ['c', 'm', 'm', ' '] => ['B', 'P', 'P', ' ']

5. get_sound_chunked() looks for sound files

• Tries biggest matching chunk (4 to 1) => not applicable here

• Fallback to three

• first cherche for BPS (Beep, Piano, Sifflement) with surprise tone

• No surprise tone => Fallback to neutral and find neutral.

• Looks for Beep Piano Piano then

• Fallback to Beep Piano neutral

5. Files are combined With pydub.AudioSegment, then played with simpleaudio

---

License

This project is community-driven and intended for non-commercial and fan-based use only.
Please verify individual file licenses before using them.

---

Thanks

Special thanks to:

• vigonotion/tts.astromech for inspiring this project!
• BD-1 Sound Database Spreadsheet for helping navigate the game’s audio files.

---

Thank you for your help! May the Force be with you.

Join the Mission

To contribute:

1. Fork this repo
2. Tweak or add functionality
3. Submit a pull request

Let’s bring BD-1 to life – together! 💬🔊✨