Model card - Kinyarwanda TTS model
Model details
- Kinyarwanda Text to Speech model
- Developed by Digital Umuganda, Arxia and Zevo Tech
- Model based from: Fastspeech and Waveglow
- License: Mozilla 2.0 License
- Feedback on the model: [email protected]
Metrics
- We use Mean Opinion Score (MOS) to evaluate the model with a maximum score being 5
| Test Corpus | MOS |
|---|---|
| Custom phrases | 3 |
Challenges
- The model does not always capture the Kinyarwanda tones
Recommendations
- Use a tonal dictionary to train future models
- Add a numbers and symbols Dictionary
- Create a code-switching dictionary containing foreign words used in Kinyarwanda
Inference API (serverless) does not yet support fastpitch models for this pipeline type.