How to speed up recognize speed?

Hello everyone~
I’m just a freshman in speech2text, and now I’ m using the pre-trained model to test some audio.
For my pc, if I recognize a 10s audio, it will cost 1s.
It’s too slow for me.
How can I speed it up?
My GPU is GTX1070 with 8G GPU memory.

Are you using the beam decoder?
If so, what beam width are you using?
You can drop the beam width to speed up. Depending on what you are using now, you may or may not see any performance drop.

Other ways to speed is to look at faster CPU. Some parts of the workload is still on CPU, so there could be bottle necks.

Another question is whether you are running things in real-time streaming 1 audio at a time, or are you running lots of audio in a batch?

Running a single audio in real time, you are already at 0.1 real time factor, which is plenty fast.

If you are running things in batch, checkout evaluate.py, it has a procedure to run predict in a batch, which will be immensely faster than using the streaming context.

Thanks~
I used client.py in native_client/python folder before. I set the beam width to 10, but nothing improved.
My work is recognize lots of audio files.
Now I am having a look at evaluate.py, little complex.

You should not, this is only intended to be used once installed from the deepspeech python wheel. trying to use it in other context might result in strange bugs.

Very thanks for your advice~