Africa TTS

Africa TTS

Speaks text aloud in 92 African languages and English, faster than real time on a CPU, with voice cloning from a short clip. For developers and researchers who need African-language speech.

Made by AfriSpeech on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Africa TTS is a text-to-speech model from AfriSpeech for 92 African languages and English. It has 110M parameters and, per the card, runs about 2.4 times faster than real time on a CPU. Every language has a default voice, and you can clone a voice from a short clip. Use only voices you have the right to use.

Who it suits: Developers, teachers and researchers who need spoken audio in African languages, where few speech tools exist. The license is non-commercial, so it suits learning, research and personal projects.

Title: Who uses it, and how
Industry | Who uses it | How they use it | Example
Education | Teachers | Make audio lessons in local languages | A teacher turns a short text in Twi into a spoken clip
Software | App developers | Add speech output to apps, including phones through an int8 bundle | A developer reads messages aloud in Swahili
Accessibility | Content creators | Provide read-aloud audio for text | A writer produces audio for an article in Hausa
Research | Speech researchers | Study low-resource language synthesis | Compare error rates against another system on the same sentences

Skip it for commercial products, because the card's license does not allow commercial use.

Where to use it:

• Python on a CPU, which is the default
• Python on an NVIDIA GPU, using the cuda device
• Phones, through the int8 bundle the card mentions
• Windows, macOS or Linux for downloading the files

Use case | Who | What they get back
Speak a sentence | Developers | A 24 kHz mono audio clip of the text
Clone a voice | Creators | Speech in a voice copied from a clip of 3 to 10 seconds
Run on a phone | Mobile developers | Speech from an int8 bundle for sherpa-onnx

How to set it up: You need Python and pip. English also needs the espeak-ng program, which the card names.

• Python and pip
• espeak-ng, if you want English
• Optionally an NVIDIA GPU

1. Install the africa-tts package.
2. Load the model in Python, which uses the CPU by default.
3. Generate speech and save it as a WAV file.

pip install africa-tts   # English also needs espeak-ng

The Hugging Face route fetches the repository files instead.

pip install -U huggingface_hub
hf download AfriSpeech/africa-tts

How to use it: Load the model, call generate with text and a language code, and save the result. Spell out numbers in your text, as the card advises.

from africa_tts import AfricaTTS

tts = AfricaTTS.load()                                                  # PyTorch CPU (default)
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")

tts_gpu = AfricaTTS.load(device="cuda")                                 # PyTorch on NVIDIA GPU

Tool | What it does
AfricaTTS.load | Loads the model on the CPU, or on a GPU with the cuda device
tts.generate | Makes audio from text, with a language code and an optional voice file
tts.save | Writes the audio to a WAV file

Example: You pass a Twi greeting with the language set to twi. The call returns audio in the default Twi voice, and saving it gives a 24 kHz mono WAV file named out.wav.

For newcomers and experts: If you are new, install the package, run the first three lines of the sample, and play the file it writes. Then try another language code.

For people in the field, the model is a 6-layer Pocket TTS FlowLM with the Mimi codec, distilled from a guided 24-layer base so the guidance sits in its weights. It reads IPA through a separate converter. Training used 897 hours of real speech. Against MMS-TTS it scored a lower character error rate in 21 of 27 shared languages.

Text and a language code go in
Text is converted to IPA
The model generates audio
Save as a 24 kHz mono file

Good to know:

Title: Character error rate for Africa TTS, lower is better
Swahili: 1.2%
English: 2.4%
Kirundi: 3%
Kinyarwanda: 3.2%
Hausa: 4.2%
Amharic: 5.7%
Yoruba: 17%
Fon: 22.5%

• License: CC BY-NC 4.0, which bars commercial use
• Cost to run: none named, you use your own CPU or GPU
• Data it reaches: it downloads files from Hugging Face, and voice cloning reads the audio clip you supply
• Not for: commercial products, or cloning a voice you have no right to use
• Quality varies by language, and the card's error rate is higher for some, such as Fon at 22.5

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: pocket-tts
• Parameters: 110M
• Weights: about 438 MB in F32; about 55 MB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 46
• Likes: 4
• License: cc-by-nc-4.0
• Languages: ada, aha, ajg, am, amh, bas, bci, bem
• Based on: kyutai/pocket-tts
• Tags: pocket-tts, onnx, safetensors, gguf, text-to-speech, african-languages, voice-cloning, multilingual

Links

• Model: https://huggingface.co/AfriSpeech/africa-tts
• Files: https://huggingface.co/AfriSpeech/africa-tts/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download AfriSpeech/africa-tts

Then follow the usage in its model card. Nothing in this kit runs by itself.