2025 · Speech
ASR lab
RNN, GRU, BiLSTM, and attention ASR on AfriSpeech-200 (Twi) and LibriSpeech, reported in WER and CER.
4recurrent model families benchmarked
- My role
- Built the training pipelines and ran the benchmarks.
- Result
- WER and CER compared across vanilla RNN, attention RNN, GRU and BiLSTM, on AfriSpeech-200 Twi and on LibriSpeech.
- Built with
- PyTorch, Torchaudio, CTC, Hugging Face
Problem
Most speech models learn from English audiobooks. How do standard recurrent architectures cope with Twi?
My role
Built the training pipelines and ran the benchmarks.
What I built
- CTC training on AfriSpeech-200 Twi
- MFCC and spectrogram feature extraction
- WER and CER across recurrent families
Outcome
WER and CER compared across vanilla RNN, attention RNN, GRU and BiLSTM, on AfriSpeech-200 Twi and on LibriSpeech.