Experiment Results

TedSrt

200 MB of audio data was scraped from Ted2Srt. We tried training on 1 GB of dataset but the results were roughly similar.

Baseline Model

Extractor	Classifier	CTC-WER	CTC-Loss
CNN	RNN	0.92	6.56

Comparing performance of Extractor

Extractor	Classifier	CTC-WER	CTC-Loss
MLP	RNN	0.99	6.58
ANN	RNN	1	6.92
RNN	RNN	0.95	6.53

Comparing performance of Classifier

Extractor	Classifier	CTC-WER	CTC-Loss
CNN	MLP	0.90	5.75
CNN	CNN	0.95	6.75
CNN	ANN	0.99	6.98

---

LibriSpeech

This is a public dataset by OpenSLR. The repository is already designed to accommodate this dataset.

Baseline Model

Extractor	Classifier	CTC-WER	CTC-Loss
CNN	RNN	0.21926	1.32976

Comparing performance of Extractor

Extractor	Classifier	CTC-WER	CTC-Loss
MLP	RNN	0.99	6.63263
ANN	RNN	0.9954	6.87
RNN	RNN	0.263	1.62

Comparing performance of Classifier

Extractor	Classifier	CTC-WER	CTC-Loss
CNN	MLP	0.5884	4.092
CNN	CNN	1	7.031
CNN	ANN	0.9982	7.065