arthemis-tts 0.1.1

A simple transformer-based text-to-speech library

Arthemis TTS

A simple and efficient transformer-based text-to-speech library for Python.

Overview

Arthemis TTS is a PyPI package that provides an easy-to-use interface for interacting with arthemis-TTS model

Features

• Simple API: Easy-to-use functions for text-to-speech conversion
• Pretrained Models: Use ready-to-go pretrained models
• GPU Support: Automatic GPU acceleration when available
• Multiple Output Formats: Support for various audio formats (WAV, MP3, etc.)
• Lightweight: Minimal dependencies and efficient implementation works on CPU

Installation

From PyPI (Recommended)

pip install arthemis-tts

Using Pretrained Models

Basic Usage

import arthemis_tts

# Simple text-to-speech with pretrained model (tested example)
model_path = "your_model.pt"
audio = arthemis_tts.text_to_speech("Hello, world!", model_path=model_path)

# Save to file
arthemis_tts.text_to_speech("Hello, world!", 
                           model_path=model_path,
                           output_path="hello_world.wav")

Advanced Usage

from arthemis_tts import load_model
from arthemis_tts.audio_processing import write_audio_to_file

# Load a pretrained model (tested with actual model)
model_path = r"C:\Users\haris\Downloads\train_ArthemisTTS (7).pt"
model = load_model(model_path)

# Generate speech
audio = model.synthesize("This is a test of the synthesize function.")

# Save the audio
write_audio_to_file(audio, "synthesized_speech.wav")

Step-by-Step Usage (Complete Example)

import torch
from arthemis_tts import ArthemisTTS
from arthemis_tts.text_processing import text_to_sequence
from arthemis_tts.audio_processing import inverse_mel_spec_to_wav, write_audio_to_file
from arthemis_tts.utils import get_device

# 1. Load your pretrained model (tested example)
model_path = "your_model.pt"
device = get_device()  # Automatically detects best device
print(f"Using device: {device}")

model = ArthemisTTS(device=device)
state_dict = torch.load(model_path, map_location=device)

# Handle different state dict formats
if isinstance(state_dict, dict) and "model" in state_dict:
    model.load_state_dict(state_dict["model"])
else:
    model.load_state_dict(state_dict)
model.eval()

# 2. Convert text to sequence
text = "Hello, world!"
text_seq = text_to_sequence(text).unsqueeze(0).to(device)

# 3. Generate mel spectrogram
with torch.no_grad():
    mel_postnet, gate_outputs = model.inference(
        text_seq,
        max_length=100,  # Shorter for faster generation
        stop_token_threshold=0.5,
        with_tqdm=True  # Show progress bar
    )

# 4. Convert to audio
audio = inverse_mel_spec_to_wav(mel_postnet.detach()[0].T)

# 5. Save audio file
write_audio_to_file(audio, "step_by_step_output.wav")
print(f"Generated audio shape: {audio.shape}")

API Reference

Main Functions

text_to_speech(text, model_path=None, output_path=None, max_length=800, gate_threshold=0.5)

Convert text to speech using a pretrained model.

Parameters:

• text (str): Input text to synthesize
• model_path (str): Path to pretrained model file (required)
• output_path (str, optional): Path to save audio file
• max_length (int): Maximum generation length (default: 800)
• gate_threshold (float): Stop token threshold (default: 0.5)

Returns:

• torch.Tensor or None: Audio tensor if no output_path, None if saved to file

load_model(model_path)

Load a pretrained model.

Parameters:

• model_path (str): Path to pretrained model file

Returns:

• ArthemisTTS: Loaded model instance

Classes

ArthemisTTS

Main TTS model class for using pretrained models.

Methods:

• inference(text_tensor, max_length=800, stop_token_threshold=0.5, with_tqdm=True): Generate mel spectrogram
• synthesize(text, max_length=800, stop_token_threshold=0.5): High-level synthesis function

Supported Audio Formats

• WAV (recommended)
• MP3

Requirements

• Python >= 3.7
• PyTorch >= 1.9.0
• torchaudio >= 0.9.0
• NumPy >= 1.19.0
• pandas >= 1.2.0
• tqdm >= 4.60.0
• pydub >= 0.25.0 (for MP3 support)

Performance Notes

• GPU Acceleration: The model will automatically use CUDA if available
• Memory Usage: Adjust max_length parameter based on available memory
• Generation Speed: Depends on text length and hardware capabilities

Model Requirements

• Models should be saved as PyTorch state dictionaries (.pt files)
• Compatible with the transformer architecture used in this library
• Models trained on the LJ Speech dataset work best for English text

Examples

Batch Processing with Pretrained Model

import arthemis_tts

# Path to your pretrained model (tested example)
model_path = r"C:\Users\haris\Downloads\train_ArthemisTTS (7).pt"

texts = [
    "Hello, world!",
    "This is Arthemis TTS.",
    "Text-to-speech synthesis."
]

for i, text in enumerate(texts):
    arthemis_tts.text_to_speech(
        text, 
        model_path=model_path,
        output_path=f"batch_output_{i+1}.wav"
    )
    print(f"Generated audio {i+1}: {text}")

Efficient Multiple Generation (Load Once, Use Many Times)

from arthemis_tts import load_model
from arthemis_tts.audio_processing import write_audio_to_file

# Load model once (tested with actual model)
model_path = r"C:\Users\haris\Downloads\train_ArthemisTTS (7).pt"
model = load_model(model_path)

texts = [
    "Hello, this is the first sentence.",
    "This is the second sentence.",
    "And this is the third sentence."
]

# Generate multiple times without reloading model
for i, text in enumerate(texts):
    audio = model.synthesize(text)
    write_audio_to_file(audio, f"efficient_output_{i+1}.wav")
    print(f"Generated efficient audio {i+1}")

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Based on the Neural Speech Synthesis with Transformer Network paper
• Inspired by the original SimpleTransformerTTS implementation
• Uses PyTorch and torchaudio for audio processing

Support

For questions and support, please open an issue on GitHub or Huggingface.

---