TigreGotico Open Voice Notebooks

Empowering the FOSS community to train, fine-tune, and deploy state-of-the-art voice models.

This repository contains a collection of Jupyter notebooks developed by TigreGotico and the Open Voice OS community. These tools are designed to democratize access to voice AI technologies, allowing developers to create datasets and train models for Text-to-Speech (TTS), Wake Word detection, and Intent Classification using open-source tools.

📂 Repository Structure

🗣️ Text-to-Speech (TTS)

Located in /tts

Tools for creating datasets and training VITS-based models.

Notebook	Description
tts_dataset_gen.ipynb	Synthetic TTS Dataset Generator. Creates LJSpeech-style datasets using a single "donor" TTS voice and Voice Conversion (VC). Features a full pipeline: synthesis, super-resolution, silence trimming, and metadata generation.
asr2tts.ipynb	ASR-to-TTS Pipeline. Converts "in-the-wild" ASR datasets (like Mozilla Common Voice) into high-quality TTS training data. Includes format standardization, denoising (resemble-enhance), silence trimming, volume normalization, and WPM filtering.
train_vits.ipynb	Train & Export VITS. A platform-agnostic notebook (Colab, Kaggle, Local) to train models using phoonnx. Supports fine-tuning, multi-speaker training, and exporting to ONNX for use with Piper, Sherpa-ONNX, and OVOS.

🔔 Wake Word (WW)

Located in /ww

Tools for generating synthetic wake word data to bootstrap training without user recordings.

Notebook	Description
tts2ww.ipynb	Wake Word Dataset Generator. A comprehensive pipeline that generates positive and negative samples. Features adversarial generation (using LLMs and grapheme edits to create similar-sounding words), TTS synthesis, voice cloning augmentation, and environmental augmentation (noise/reverb) for robust model training.

🧠 Intent Classification (M2V)

Located in /m2v

Efficient, multilingual intent recognition for offline voice assistants.

Notebook	Description
ovos_intent_classifier_multilingual.ipynb	Multilingual Intent Classifier. Trains extremely efficient classifiers using model2vec on the Open Voice OS intents dataset. Includes steps to export the model to ONNX for dependency-free inference (requiring only numpy and onnxruntime).

📝 Text Utilities

Located in /arabic_diacritics

Notebook	Description
lstm.ipynb	Arabic Diacritizer. Trains a lightweight LSTM model to automatically add diacritics to Arabic text. This is a critical preprocessing step for training high-quality Arabic TTS models. Includes export to ONNX.

---

🚀 Getting Started

These notebooks are designed to be self-contained. Most define their own dependencies and installation steps within the first few cells.

Prerequisites:

1. Python 3.10+ (Recommended).
2. GPU Acceleration: While inference steps can run on CPU, training (VITS) and heavy data processing (Voice Conversion/Denoising) are significantly faster with an NVIDIA GPU (CUDA).
3. HuggingFace Account: Some notebooks require a token to upload datasets or download gated models.

Usage:

1. Clone this repository:

   git clone <your-repo-url>
   cd <your-repo-name>

2. Launch Jupyter Lab or Notebook:

   jupyter lab

3. Open the desired notebook and follow the "Configuration" cells at the top of each file to set your paths and parameters.

---

🤝 Community & Support

These tools are built to support the Open Voice OS ecosystem and the broader privacy-focused AI community.

• Open Voice OS: openvoiceos.org
• Matrix Chat: #openvoiceos:matrix.org

📜 Credits & Acknowledgments

DISCLAIMER: Generative AI was used to convert various python scripts into notebook format

• Author: TigreGotico
• Funding: Funded through the NGI0 Commons Fund via NLnet, with support from the European Commission's Next Generation Internet programme (grant No 101135429).
• Core Technologies:
  • Phoonnx / VITS
  • chatterbox-onnx
  • Model2Vec
  • ONNX Runtime

⚖️ License

Apache 2.0 (or see individual notebooks for specific licensing details).