I have trained a speech recognition model which uses CTCLoss and is inspired from


I trained it on the Librispeech Dataset (train-clean-100) comprising 100 hrs of recordings. The model performs decently well on the LibriSpeech test dataset (test-clean). For example for one of the samples from the test dataset, it gives output like below;-

Ground Truth - he hoped there would be stew for dinner turnips and carrots and bruised potatoes and fat mutton pieces to be ladled out in thick peppered flour fattened sauce.

Model Output - he hoped there would be sto her dinner turnips and charats and brused betta ose and fat bmutten pieces to be latled out ind thick peppered flouer faton sause.

However, when I am using my own recording-which I record from my IPhone XS Max, convert it to .wav and load it using torchaudio-I get a completely gibberish output. For example my recording is "what are you having for lunch?" it gives output like "oui" which is nowhere related to the original recording. I have made sure that the recording is loud and clear with no background noise. I do the same data processing ( converting to mel spectrograms) that i did on the original dataset.

Am I missing something here? I am a bit new to this so any help would be appreciated.

  • $\begingroup$ Speech recognition in non-academic conditions is tremendously hard. You should either pre-process your recording to have similar properties asLibrispeech (the same energy, remove noise, etc.) or modify the training data to look more like a noisy phone recording. $\endgroup$ – Jindřich Nov 24 '20 at 9:10
  • $\begingroup$ Hi thanks, all of the librispeech recordings are in a .flac format, but the one from my iphone is in .wav. Can the formats cause an issue? I am using torchaudio for all preprocessing. $\endgroup$ – Divyaanand Sinha Nov 25 '20 at 14:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.