Africa TTS

Text-to-speech for 92 African languages and English, with voice cloning, built on Kyutai's Pocket TTS. Every language has a default voice, so all you need is text and a language code.

🎙️ Try it live · 🔊 Samples · 🤗 Model

An ongoing project: native-speaker feedback is what improves it most. See CONTRIBUTING.md.

⚡ Runs on a CPU, no GPU needed

Africa TTS runs on audio.cpp by default: a 144 MB model, no PyTorch, faster than real time on an ordinary laptop, streaming its first audio about 0.1 s after it gets the text. On a 2020 laptop CPU (AMD Ryzen 5 4500U, 4 threads) it speaks 2.3-2.8× faster than real time. It has 110M parameters, so it runs on laptops, servers and phones.

The unquantised PyTorch engine is optional (pip install "africa-tts[torch]"): it also runs on a GPU, and on a 4-vCPU cloud machine it speaks 2.4× faster than real time (speed benchmark).

🎯 Intelligible

On held-out sentences transcribed by Omnilingual ASR, Africa TTS is more intelligible than MMS-TTS in 21 of 27 languages. The 813M OmniVoice, seven times its size, is more intelligible in 24 of 29, by small margins: benchmark.

Install

pip install africa-tts                 # default: audio.cpp on the CPU, no PyTorch
pip install "africa-tts[torch]"        # optional: the unquantised PyTorch engine (CPU or GPU)

English also needs espeak-ng (apt install espeak-ng / brew install espeak-ng). The first time, it asks before downloading the model (about 150 MB; 440 MB for PyTorch) from Hugging Face, then caches it and works offline. In scripts, approve the download with --yes, AfricaTTS.load(download=True), or africa-tts --download beforehand. For PyTorch on a CPU-only Linux machine, add --extra-index-url https://download.pytorch.org/whl/cpu to get the small CPU build.

Use

africa-tts "Akwaaba! Yɛfrɛ me Kofi." --lang twi -o twi.wav
africa-tts --list-languages
africa-tts serve                       # the API server below, on port 8000

from africa_tts import AfricaTTS

tts = AfricaTTS.load()                                                  # audio.cpp, CPU
audio = tts.generate("Akwaaba! Yɛfrɛ me Kofi.", lang="twi")              # default Twi voice
audio = tts.generate("Hello Ghana!", lang="eng", voice="my_voice.wav")  # clone a 3-10 s clip
tts.save(audio, "out.wav")

for chunk in tts.stream("Akwaaba! Yɛfrɛ me Kofi.", lang="twi"):         # 80 ms chunks
    ...

tts = AfricaTTS.load(backend="torch")   # PyTorch, unquantised (pip install "africa-tts[torch]")
tts = AfricaTTS.load(device="cuda")     # PyTorch on an NVIDIA GPU

• Languages: ISO 639-3 codes (twi, hau, yor, swh, zul, amh, eng, ...); see africa-tts --list-languages.
• Numbers: spell them out ("twenty", not "20").

Run it elsewhere

More

• Benchmarks: speed, and intelligibility per language
• Model, how it was made, data and limitations
• Training recipe · ONNX export · training code
• Status and roadmap · Contributing

License

Code: MIT (LICENSE). Model and default voices: CC-BY-NC-4.0. espeak-ng (English) is GPL-3.0-or-later. Built by AfriSpeech.