Model card - Kinyarwanda Deepspeech model
Model details
- Kinyarwanda Speech to text model
- Developed by Mozilla and Digital Umuganda
- Model based from: Baidu Deepspeech end to end RNN model
- paper: deepspeech end to end STT
- Documentation on model: deepspeech documentation
- License: Mozilla 2.0 License
- Feedback on the model: joshua.richard.meyer@gmail.com and samuelrutunda@digitalumuganda.com
Intended use cases
- Intended to be used for
- simple keyword spotting
- simple transcribing
- transfer learning for better kinyarwanda and african language models
- Intended to be used by:
- App developpers
- various organizations who wants to transcribe kinyarwanda recordings
- ML researchers
- other researchers in Kinyarwanda and tech usage in kinyarwanda (e.g. Linguists, journalists)
- Not intended to be used as:
- a fully fledged voice assistant
- voice recognition application
- Multiple languages STT
- language detection
Factors
- Anti-bias: these are bias that can influence the accuracy of the model
- Gender
- accents and dialects
- age
- Voice quality: factors that can influence the accuracy of the model
- Background noise
- short sentences
- Voice format: voices must be converted to the wav format
- wav format
Metrics
- word error rate on the Common Voice Kinyarwanda test set
| Test Corpus | WER | CER |
|---|---|---|
| Common Voice | 60.1% | 23.5% |
Training data
Evaluation data
Caveats and recommendation
- More accents other than main kinyarwanda accents must be included
- Language model to correct grammatical errors needed
- More unique individual voice must be included in the datasets
- Smaller models that can be incorporate in mobile devices
Quantitative data
- coming soon...