I'm training a Mozilla's DeepSpeech model using public domain data, I'll be releasing the data I collect also under public domain.
Automatically aligned the text with the Windows speech recognition, then as validation of the alignment used a Mozilla's DeepSpeech model using a few different language models. The first model to start the validation was trained on voxforge Spanish data and on top of the ones that scored the highest confidence with the Windows Speech.
Initially I'm collecting only Spanish, eventually I'll add more languages.
- Spanish 120h of clean speech
100h of clean speech from a single speaker LJSpeech format!
Please if you find wrong transcription open an issue or send a PR to fix them.
Public domain.