search

SeanNaren / deepspeech.torch

Speech Recognition using DeepSpeech2 network and the CTC activation function.
MIT License
259 stars 73 forks source link

Issue with Torch Dependencies #88

Open nn-learner opened 7 years ago

nn-learner commented 7 years ago

Hi Sean,

I don't know whether you are still working on maintaining this repo. However, I recently had to upgrade environments and it removed my torch libraries. I'm have an issue with the dependencies. It seems cunnx is no longer supported by the latest version of torch. I was wondering if you knew of a work around.

SeanNaren commented 7 years ago

I don't think cunnx is used in this repo, could you point me to where it is?

nn-learner commented 7 years ago

It says on the wiki for the required dependencies to compile Baidu's CTC.

SeanNaren commented 7 years ago

Thanks for that, it should not be necessary (but nnx is)! I've removed that from the install at the wiki page.

nn-learner commented 7 years ago

Interesting so when I try to transcribe the audio sample with one of my pre-trained weights, I get the following output:

th Predict.lua -audioPath audio_samples/1089-134691-0009.flac mspsbsfsspsfsmsmspsysnsmsmspsmspsmsmpbgbysssmsysymsmskdspmbmbspkspdbsksmydysmsgpysklbddsbd

I have been using this pretrained weight before in my other environment and it transcribed perfectly fine. Any suggestions why this might be occuring?

Also, I have been moving these torch dependencies to a docker, which I can share with you if you want to add it to your repo documentation.

SeanNaren commented 7 years ago

Is the dictionary correct? Seems like something's up with the decoding

nn-learner commented 7 years ago

Yes, I tried both on a fresh git pull from your repo and my existing instance of the code I pulled.