Reputation: 1
I tried the following code:
#!/usr/bin/env python3
import sys
import json
from vosk import BatchModel, BatchRecognizer, GpuInit
from timeit import default_timer as timer
TOT_SAMPLES = 0
GpuInit()
model = BatchModel("model")
with open(sys.argv[1]) as fn:
fnames = fn.readlines()
fds = [open(x.strip(), "rb") for x in fnames]
uids = [fname.strip().split("/")[-1][:-4] for fname in fnames]
recs = [BatchRecognizer(model, 16000) for x in fnames]
results = [""] * len(fnames)
ended = set()
start_time = timer()
while True:
# Feed in the data
for i, fd in enumerate(fds):
if i in ended:
continue
data = fd.read(8000)
if len(data) == 0:
recs[i].FinishStream()
ended.add(i)
continue
recs[i].AcceptWaveform(data)
TOT_SAMPLES += len(data)
# Wait for results from CUDA
model.Wait()
# Retrieve and add results
for i, fd in enumerate(fds):
res = recs[i].Result()
if len(res) != 0:
results[i] = results[i] + " " + json.loads(res)["text"]
if len(ended) == len(fds):
break
end_time = timer()
for i, res in enumerate(results):
print(uids[i], res.strip())
print("Processed %.3f seconds of audio in %.3f seconds (%.3f xRT)"
% (TOT_SAMPLES / 16000.0 / 2,
end_time - start_time,
(TOT_SAMPLES / 16000.0 / 2 / (end_time - start_time))),
file=sys.stderr)
source: https://github.com/alphacep/vosk-api/blob/master/python/example/test_gpu_batch.py
but got:
C:\Users\user\PycharmProjects\vosk-test\.venv\Scripts\python.exe C:\Users\user\PycharmProjects\vosk-test\main.py
Traceback (most recent call last):
File "C:\Users\user\PycharmProjects\vosk-test\main.py", line 16, in <module>
model = BatchModel("vosk-model-en-us-0.42-gigaspeech")
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\user\PycharmProjects\vosk-test\.venv\Lib\site-packages\vosk\__init__.py", line 243, in __init__
raise Exception("Failed to create a model")
Exception: Failed to create a model
Process finished with exit code 1
How to run Vosk on GPU using python?
Upvotes: 0
Views: 442
Reputation: 1
For me, simply recompiling kaldi and vosk like in the Dockerfile of Vladimir link did the trick, don’t use the gpu fork branch but the main vosk.
Upvotes: 0