Animalese TTS

Generate Animalese speech (the cute gibberish language from Animal Crossing) from any text input.

Features

• Clean Synthesizer - Rule-based Animalese generator with crystal-clear audio
• Multiple Voice Types - High (peppy), normal, and low (cranky) voices
• Adjustable Speed - Control how fast characters speak
• Neural TTS Model - PyTorch-based model for training on custom data (experimental)

Quick Start

Installation

# Create conda environment
conda create -n animalese-tts python=3.10 -y
conda activate animalese-tts

# Install dependencies
pip install -r requirements.txt

Generate Animalese Audio

# Basic usage
python animalese_clean.py --text "Hello, welcome to my island!" --output hello.wav

# Different voice types
python animalese_clean.py --text "Your text" --voice high --output cute.wav    # Peppy villager
python animalese_clean.py --text "Your text" --voice normal --output normal.wav
python animalese_clean.py --text "Your text" --voice low --output grumpy.wav   # Cranky villager

# Adjust speed (lower = faster)
python animalese_clean.py --text "Excited!" --speed 45 --output fast.wav
python animalese_clean.py --text "Slow speech" --speed 70 --output slow.wav

# Adjust base pitch
python animalese_clean.py --text "Higher pitch" --pitch 300 --output high_pitch.wav

Synthesizer Options

Option	Description	Default
--text	Text to convert to Animalese	(required)
--output	Output WAV file path	animalese_clean.wav
--voice	Voice type: high, normal, low	normal
--pitch	Base pitch in Hz	240.0
--speed	Duration per character in ms (lower = faster)	55

Project Structure

animalese-tts/
├── animalese_clean.py      # Main synthesizer (recommended)
├── animalese_v2.py         # Alternative synthesizer with more features
├── simple_animalese.py     # Basic synthesizer
├── train.py                # Neural TTS training script
├── inference.py            # Neural TTS inference script
├── configs/
│   ├── default.yaml        # Full training config
│   └── demo.yaml           # Quick demo config
├── models/                 # Neural network components
│   ├── animalese_tts.py    # Main TTS model
│   ├── encoder.py          # Character encoder
│   ├── decoder.py          # Mel decoder
│   ├── vocoder.py          # HiFi-GAN vocoder
│   └── ...
├── data/                   # Dataset utilities
├── scripts/                # Data generation scripts
└── utils/                  # Training utilities

Neural TTS (Experimental)

The project also includes a neural TTS model based on FastSpeech2 + HiFi-GAN architecture. This requires training on Animalese audio samples.

Generate Training Data

python scripts/generate_synthetic_data.py --output-dir data/train --num-samples 1000

Train the Model

python train.py --config configs/default.yaml --data-dir data/train

Run Inference

python inference.py --checkpoint checkpoints/best.pt --text "Hello!" --output output.wav

Note: The neural model requires significant training time (hundreds of epochs) and ideally real Animalese samples to produce good results. For quick results, use animalese_clean.py.

How It Works

Clean Synthesizer (animalese_clean.py)

The rule-based synthesizer creates Animalese by:

1. Character Mapping - Each letter maps to a vowel-like sound
2. Formant Synthesis - Uses vocal formant frequencies for natural speech quality
3. Pitch Variation - Varies pitch based on character and position for melodic speech
4. Smooth Envelopes - Attack-sustain-release for each phoneme

This mimics how the actual Animal Crossing games generate Animalese - each character produces a short vocal sound with pitch variation.

Neural TTS

The neural approach uses:

• Character Encoder - Transformer-based encoding of input text
• Variance Adaptor - Predicts duration, pitch, and energy
• Mel Decoder - Generates mel spectrograms
• HiFi-GAN Vocoder - Converts mel spectrograms to audio

Real-Time Speaking

Speak text immediately without saving to a file:

# Speak text directly
./speak.py "Hello! I can speak Animalese!"

# Pipe text to speak
echo "This is amazing!" | ./speak.py --stdin

# Read from file and speak
cat story.txt | ./speak.py --stdin

# Different voices
./speak.py "I'm peppy!" --voice high
./speak.py "I'm cranky." --voice low

# Adjust speed
./speak.py "Fast!" --speed 2.0      # 2x faster
./speak.py "Slow..." --speed 0.5    # Half speed

Make Claude Code speak to you

You can pipe my responses to the speaker:

# In your terminal, create a wrapper
echo 'Hello from Claude!' | ./speak.py --stdin

# Or integrate into your workflow
claude "tell me a joke" | tee /dev/tty | ./speak.py --stdin

Examples

# Villager greeting
python animalese_clean.py --text "Hello! How are you today?" --voice high --output greeting.wav

# Tom Nook style
python animalese_clean.py --text "Yes, yes! That will be fifty thousand bells." --voice normal --pitch 200 --output nook.wav

# Excited discovery
python animalese_clean.py --text "Oh wow! I found a rare fossil!" --voice high --speed 45 --output excited.wav

# Long dialogue
python animalese_clean.py --text "Let me tell you about my day..." --voice low --speed 60 --output story.wav

# Real-time speaking
./speak.py "I can speak in real-time now!"

License

This project is for educational and research purposes.

Acknowledgments

• Animalese is a trademark of Nintendo
• Neural architecture inspired by FastSpeech2 and HiFi-GAN