detect_language() doesn't work on a GPU?

[utility-aagrawal]

Here's a simple script to identify language from an audio:

import whisperx
import time

start_time = time.time()
filepath = ""

whisper_model = whisperx.load_model("medium", device = "cuda", compute_type="float16")

audio = whisperx.load_audio(filepath)
audio = whisperx.audio.pad_or_trim(audio)

print(f"Language => {whisper_model.detect_language(audio)}")

end_time = time.time()
print(f"Time taken: {end_time - start_time:.2f} seconds")

You can use this file as an input:

Download-30secs.mp4

I am getting this error:

Could not load library libcudnn_cnn_infer.so.8. Error: libcudnn_cnn_infer.so.8: cannot open shared object file: No such file or directory
Please make sure libcudnn_cnn_infer.so.8 is in your library path!
Aborted (core dumped)

Can someone tell me why? Let me know if you need anythind additional. Thanks!

[utility-aagrawal]

What I don't understand is how transcribe() works even if I keep everything else in the code unchanged.

This code works:

import whisperx
import time

start_time = time.time()
filepath = ""

whisper_model = whisperx.load_model("medium", device = "cuda", compute_type="float16")

audio = whisperx.load_audio(filepath)
audio = whisperx.audio.pad_or_trim(audio)

results = whisper_model.transcribe(audio)
print(results)

end_time = time.time()
print(f"Time taken: {end_time - start_time:.2f} seconds")

transcribe() also uses the same detect_language() method as you can see here:

whisperX/whisperx/asr.py

Line 194 in f2da2f8

language = language or self.detect_language(audio)

but this doesn't throw the same error.

[candrascik]

Adding the following to my Dockerfile fixed this issue. Make sure nvidia-cudnn-cu12 is <9.

RUN pip install nvidia-cudnn-cu12==8.9.7.29
ENV LD_LIBRARY_PATH /usr/local/lib/python3.10/dist-packages/nvidia/cudnn/lib:$LD_LIBRARY_PATH