AfriSpeech/akiti-tts

natural sounding cpu friendly voices for african languages # Akiti TTS Text-to-speech you can run **locally on your own CPU**. Akiti TTS currently supports **Asante Twi**, with more languages planned. It is fine-tuned from VieNeu-TTS-0.3B and served as a small quantized GGUF model, so it runs without a GPU. - 🧠 Model: afrispeech/Akiti-TTS - 🌐 Web demo: afrispeech/Akiti-TTS Space The model weights download automatically from HuggingFace on first run and are cached locally afterward. ## Install **1. System dependency — espeak-ng** ```bash # Linux sudo apt-get install espeak-ng # macOS brew install espeak-ng # Windows: github.com ``` **2. Python packages** ```bash pip install -r requirements.txt ``` > Note: `llama-cpp-python` may build from source on first install (a few > minutes). Prebuilt CPU wheels are available for some platforms. ## Usage ```bash # Synthesize to a WAV file python akiti_local.py --text "Meda wo ase paa." --output out.wav # Higher quality (Q8) model python akiti_local.py --text "Meda wo ase." --model q8 --output out.wav ``` ### Studio (web app) For creating voice content, run the browser-based **Akiti TTS Studio**: ```bash python akiti_web.py ``` It starts a small local server and opens `127.0.0.1` in your browser. The model loads once and stays warm, so each generation is fast. From there you can: - write or paste a script and **Generate** (or press **Ctrl/⌘ + Enter**), - play clips inline and **Download** them as WAV, - build up a session **library** of generated clips, - tune the model (q4/q8), temperature, and repetition penalty. Options: `--model q8`, `--port 8000`, `--host 0.0.0.0`, `--no-browser`. Everything runs locally — nothing is uploaded. ### REST API server To call Akiti TTS from your own code or tools (curl, Python, n8n, etc.), run the REST API server: ```bash python akiti_api.py # loads q8, serves on 127.0.0.1 python akiti_api.py --model q4 --port 9000 ``` The model loads once and stays w …