音声
hyprwhspr
Native Linux dictation: fast, private, and hackable speech-to-text.
★ 1.2k スターMIT
ガイドを見る →IMS-Toucanは、DigitalPhoneticsが公開しているオープンソースの音声プロジェクトです（GitHub: DigitalPhonetics/IMS-Toucan）。GitHubスターは2,208、フォークは318で、Apache-2.0ライセンスの下で公開されています。GitHubでの最終更新は2026年1月25日です。
ToucanTTS: fast, controllable text-to-speech for over 7,000 languages, from Stuttgart's IMS.
Decide whether you need speech in one of the pretrained model's languages, or a fine-tuned model for your own voice or language data.
Clone the repository and install system packages such as libsndfile1, espeak-ng, and ffmpeg, then create a virtual environment and install the Python requirements.
For custom work, prepare a training recipe by registering your dataset's audio paths and transcripts and copying a fine-tuning example.
Run inference through the ToucanTTS interface or the provided reader scripts, or launch a training pipeline run for your recipe.
Pretrained ToucanTTS models cover over 7,000 languages, with a published list of supported languages and a free hosted demo on Hugging Face.
Scaling parameters let you adjust duration, pitch-curve variance, and energy-curve variance at inference time.
Models download automatically when needed, and the interface supports setting language and speaker embeddings.
Example pipelines for single-dataset and multilingual fine-tuning show how to adapt the model to new data with modest compute.
Train from scratch or fine-tune with the training pipeline script, including multi-GPU runs via torchrun and optional Weights and Biases logging.
A scorer utility can find and remove outlier samples with unusually high loss from a dataset cache before training.
Install the listed system packages on Linux first, then set up the virtual environment.
git clone https://github.com/DigitalPhonetics/IMS-Toucan.git
cd IMS-Toucan
pip install --no-cache-dir -r requirements.txtThe text-to-file reader script synthesises a list of sentences into one audio file; pretrained weights download automatically.
python run_text_to_file_reader.py # see README inference section for interface optionsAfter preparing a recipe, launch the training pipeline; add --finetune to start from a checkpoint.
python run_training_pipeline.py --finetuneIt is developed and maintained by researchers at the Institute for Natural Language Processing, University of Stuttgart, published under DigitalPhonetics.
Yes. The code and models are Apache-2.0 licensed and free, and the hosted Hugging Face demo is free to try.
No API key is needed; models run locally in PyTorch and download from Hugging Face automatically.
Commercial TTS APIs cover far fewer languages and charge per character; Toucan focuses on extreme multilingual coverage and trainability, at the cost of running and managing the model yourself.
Clone the repo, install the system packages and requirements.txt, then run the text-to-file reader script; the pretrained model downloads on first use.
Native Linux dictation: fast, private, and hackable speech-to-text.
Fast and local neural text-to-speech engine
Fast local speech-to-text across 16+ model families, powered by ggml.
IndexTTS: industrial-level, controllable zero-shot text-to-speech with voice cloning
eSpeak NG — the compact open-source speech synthesizer for 100+ languages, small enough to embed anywhere.
MaryTTS — open-source, multilingual text-to-speech platform written in pure Java
LibreAIカタログに掲載された音声プロジェクトの中で最もスターの多いもの — IMS-Toucanが合わなかった場合に検討したい、現実的な候補リストです。
利用 AI 大模型和自动化工作流，根据主题或关键词一键生成高清短视频。Generate HD short videos from a topic or keyword with an automated AI workflow.
Robust Speech Recognition via Large-Scale Weak Supervision
1 min voice data can also be used to train a good TTS model! (few shot voice cloning)
Clone a voice from 5 seconds of audio and generate arbitrary speech in real time
Port of OpenAI's Whisper model in C/C++
カタログの中でIMS-Toucanとスター数が近い音声プロジェクト — 前後の棚もそのまま見て回れます。
GAN vocoder that turns mel-spectrograms into high-fidelity speech audio, fast enough for real-time use.
WaveRNN vocoder plus Tacotron TTS in PyTorch: train your own neural speech synthesis on LJSpeech.
VoiceMem — streaming dual-brain memory for voice agents: a left brain for facts, a right brain for emotion
Fast local speech-to-text across 16+ model families, powered by ggml.
Offline text-to-speech for Python — uses your operating system's own voices, no internet needed
Turn books and documents into natural speech with the Kokoro TTS model.