---
title: Ghana Speech Nano - Multilingual Matcha TTS
emoji: 🍵
colorFrom: green
colorTo: yellow
sdk: gradio
app_file: app.py
pinned: false
license: mit
---

# 🍵 Ghana Speech Nano — Multilingual Matcha-TTS

Inference demo for [ghananlpcommunity/ghana-speech-nano](https://huggingface.co/ghananlpcommunity/ghana-speech-nano),
a single [Matcha-TTS](https://github.com/shivammehta25/Matcha-TTS) acoustic model finetuned across
**42 languages** from `ghananlpcommunity/ghana-speech`, using Matcha's speaker-conditioning slot as a
language ID (`n_spks=42`).

- Text is phonemized on the fly with the same espeak-ng `lfn` voice pipeline used at training time
  (`matcha.text.cleaners.twi_cleaners`), then filtered to the model's symbol set.
- Audio is vocoded with the universal HiFi-GAN checkpoint (`hifigan_univ_v1`).
- Language names are recovered dynamically from the `ghananlpcommunity/ghana-speech` dataset's
  config list (sorted the same way the training script assigned language IDs), with a generic
  fallback (`Language 0..41`) if the dataset can't be reached.
- Checkpoint files are pulled straight from the model repo's `checkpoints/` folder; use the
  "Refresh checkpoint list" button to pick up newly-pushed epochs.

Code (this fork, pinned commit `fa42c91`): https://github.com/michsethowusu/Matcha-TTS