CaroTTS-60M-DE-Caro

CaroTTS-60M-DE-Caro 🥕

<p align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/652a6ea890e317f2436fb7e1/xqaeRYV93A-NwfMv5eEft.png" width="250" alt="image" /> </p>

Fast, lightweight German Text-to-Speech model based on FastPitch + HiFi-GAN architecture. Full training and export code available at CaroTTS GitHub repository.

Model Description

This model provides high-quality German text-to-speech synthesis using a non-autoregressive architecture optimized for fast inference on CPUs and mobile devices. The model consists of two components:

• FastPitch: Duration predictor and mel-spectrogram generator (~46M parameters)
• HiFi-GAN: Neural vocoder that converts spectrograms to audio (~14M parameters)
• Voice: Single Speaker, Female Voice, 'Caro'

Total Parameters: ~60M

Key Features

• 🚀 Fast Inference: Non-autoregressive architecture enables real-time synthesis
• 💻 CPU-Friendly: Optimized for deployment on resource-constrained devices
• 🎯 High Quality: Natural-sounding German speech
• 📦 Multiple Formats: Available in Nemo, ONNX and PT2 (PyTorch Inductor) formats

Try It Out

👉 Interactive Demo on HuggingFace Spaces (uses PT2 format with Zero GPU)

Model Files

• Caro_fastpitch.nemo - FastPitch model in NEMO format
• Caro_hifigan.nemo - HiFi-GAN vocoder in NEMO format
• Caro_fastpitch.onnx - FastPitch model in ONNX format
• Caro_hifigan.onnx - HiFi-GAN vocoder in ONNX format
• Caro_fastpitch_encoder.pt2 - FastPitch-Encoder compiled with PyTorch Inductor (for CUDA/Zero GPU)
• Caro_fastpitch_decoder.pt2 - FastPitch-Decoder compiled with PyTorch Inductor (for CUDA/Zero GPU)
• Caro_hifigan.pt2 - HiFi-GAN compiled with PyTorch Inductor (for CUDA/Zero GPU)

Note: PT2 files have been exported and compiled on ZERO GPU and may have to be reexported for use on other hardware. Visit the CaroTTS GitHub repository for the export code.

Usage

ONNX Inference

ONNX provides the best compatibility and performance for CPU deployment:

import numpy as np
import onnxruntime as ort
import soundfile as sf

# Tokenization function
def normalize_unicode_text(text: str) -> str:
    import unicodedata
    if not unicodedata.is_normalized("NFC", text):
        text = unicodedata.normalize("NFC", text)
    return text

def any_locale_text_preprocessing(text: str) -> str:
    res = []
    for c in normalize_unicode_text(text):
        if c in ["'"]:
            res.append("'")
        else:
            res.append(c)
    return "".join(res)

def tokenize_german(text: str, punct: bool = True, apostrophe: bool = True,
                   pad_with_space: bool = True) -> list[int]:
    """Tokenize German text into a list of integer token IDs."""

    _CHARSET_STR = "ABCDEFGHIJKLMNOPQRSTUVWXYZÄÖÜẞabcdefghijklmnopqrstuvwxyzäöüß"
    _PUNCT_LIST = [
        "!", '"', "(", ")", ",", "-", ".", "/", ":", ";", "?", "[", "]",
        "{", "}", "«", "»", "‒", "–", "—", "'", "‚", '"', "„", "‹", "›",
    ]

    tokens = [" "]  # Space at index 0
    tokens.extend(_CHARSET_STR)
    if apostrophe:
        tokens.append("'")
    if punct:
        tokens.extend(_PUNCT_LIST)
    tokens.extend(["<pad>", "<blank>", "<oov>"])

    token2id = {token: i for i, token in enumerate(tokens)}
    space = " "

    text = any_locale_text_preprocessing(text)

    # Encode
    cs = []
    tokens_set = set(tokens)

    for c in text:
        if ((c == space and len(cs) > 0 and cs[-1] != space) or
            ((c.isalnum() or c == "'") and c in tokens_set) or
            (c in _PUNCT_LIST and punct)):
            cs.append(c)

    if cs:
        while cs and cs[-1] == space:
            cs.pop()

    if pad_with_space:
        cs = [space] + cs + [space]

    return [token2id[p] for p in cs]

# Load ONNX models
fastpitch_session = ort.InferenceSession("Caro_fastpitch.onnx")
hifigan_session = ort.InferenceSession("Caro_hifigan.onnx")

# Prepare text
text = "Hallo, ich bin CaroTTS, ein deutsches Text-zu-Sprache-System."
tokens = tokenize_german(text)

# Prepare inputs
paces = np.ones(len(tokens), dtype=np.float32)
pitches = np.zeros(len(tokens), dtype=np.float32)

inputs = {
    "text": np.array([tokens], dtype=np.int64),
    "pace": np.array([paces], dtype=np.float32),
    "pitch": np.array([pitches], dtype=np.float32),
}

# Generate spectrogram
spec = fastpitch_session.run(None, inputs)[0]

# Generate audio
audio = hifigan_session.run(None, {"spec": spec})[0]

# Save audio (44.1kHz sample rate)
sf.write("output.wav", audio.squeeze(), 44100)

NeMo Inference

If you have NeMo installed (pip install nemo-toolkit[tts])and want to work with the original .nemo checkpoints:

import torch
import soundfile as sf
from nemo.collections.tts.models.fastpitch import FastPitchModel
from nemo.collections.tts.models.hifigan import HifiGanModel

# Load models
device = "cuda" if torch.cuda.is_available() else "cpu"
fastpitch = FastPitchModel.restore_from("Caro_fastpitch.nemo", map_location=device).eval()
hifigan = HifiGanModel.restore_from("Caro_hifigan.nemo", map_location=device).eval()

# Prepare text
text = "Guten Tag. Herzlich Willkommen zu dieser Demonstration."

with torch.inference_mode():
    # Parse and generate
    parsed_text = fastpitch.parse(text)
    spec = fastpitch.generate_spectrogram(tokens=parsed_text)
    audio = hifigan.convert_spectrogram_to_audio(spec=spec)

    # Save audio (44.1kHz sample rate)
    sf.write("output.wav", audio.squeeze().cpu().numpy(), 44100)

Citation

@misc{carotts2024,
  title={CaroTTS: Fast Lightweight German Text-to-Speech},
  author={Holtzwart, Tassilo},
  year={2024},
  url={https://github.com/TassiloHo/CaroTTS}
}

Acknowledgments

• Built with NVIDIA NeMo
• FastPitch architecture: arXiv:2006.06873
• HiFi-GAN: arXiv:2010.05646

License

MIT License

---

For more information, visit the CaroTTS GitHub repository.