¿How can i solve a failure training for speech to text service?
I am starting to work with speech to text service. In test environment it works perfectly, however the production environment is very noisy so I need to create an acoustic model. The creation of the model was without problems, I added files with a total of more than 27 minutes of audio but when I start the training, after several minutes it fails with the following response error:
{
"owner": "c9e7d552-49ab-430a-a4e2-d75d6d99d9a7",
"base_model_name": "es-ES_BroadbandModel",
"customization_id": "faea5379-4f54-4a5f-a73a-7028522b3f01",
"versions": [
"es-ES_BroadbandModel.v2018-02-07"
],
"created": "2018-07-17T13:17:25.442Z",
"name": "test 1",
"description": "descripcion test 1",
"progress": 0,
"language": "es-ES",
"error": "Unresponsive backend detected. Please try later.",
"status": "failed"
}
How can I solve this error in order to add my noisy acoustic model?
Hi Arthur, I am getting the same error when I try to train the acoustic model with a trained language model. Have tried several times and it still won't work. can you please confirm if this bug has been resolved or not?
Thanks
It was a bug in our service having to do with training on stereo audio. It will be fixed in the next update. Until then, please downmix your audio before submitting it for training.
Thank you Arthur. what do you mean with downmix? switch from stereo to mono? I can changue te audio for sure, but can you share with me the optim specs? regards!
yes, please mix stereo to mono. Normally the service would do this itself, but there was a bug in it. Ideal audio is 16khz (for broadband model), mono.
Hi Arthur. Can you confirm that the bug still exists? I am getting the same response when training a custom audio model. Thanks. Aaron