Getting very poor predictions from a trained model - 0.7 version over the existing checkpoint

DeepSpeech
#1

I am trying to train english voice over deepspeech’s existing checkpoint and scorer. I have made chunks on silence for all the files that I am giving it for training. But the model’s result is coming very poor. I have run for 70 epocs with 343 train files, 35 validation and 30 test files each roughly 6 seconds.

The output is very gibberish. Can you please help? Do I need to increase the train files? Because what I assume is if I train over the pre trained checkpoint, then even the 343 train files should give me decent results when I have run for 70 epocs.

#2

More data, general models have 300,000 plus files for training.

Try transfer learning on the released model.

#3

Hi Sir,

I am using transfer learning on the existing deepspeech 0.7.0 checkpoint, but the updated model is giving very poor result. Can you please help. I am taking mostly the default parameters only for training the .wav files.

The bit rate is 256 kbps and the Frame rate is 16000. Hope I am using the right parameters. My loss values while training are hovering between 50 and 100.

Can you please provide your valuable inputs?

#4

Search the forum for “transfer learning”, but 300 files don’t get you anywhere. Use thousands.