Run Batched Faster Whisper with proper arguments

[toanhn]

After using and testing Batched Faster Whisper, I think this is a case with the proper arguments to get a good resulting transcript.
Note: Do not use "initial_prompt" because it will reduce the transcription accuracy, we should fine-tune Faster Whisper with proper nouns and terminologies.

import time

from faster_whisper import BatchedInferencePipeline, WhisperModel

start_time = time.time()

# Model configurations
model_size = "large-v3"
device = "cuda"
compute_type = "float16"

# Transcribing configurations
without_timestamps = True
word_timestamps = True
batch_size = 16
beam_size = 5
patience = 2
task = "transcribe"

audio_file = "test.wav"

start_time_load_model = time.time()

model = WhisperModel(
    model_size_or_path=model_size,
    device=device,
    compute_type=compute_type,
)

print("\nLoad model time:", time.time() - start_time_load_model, "\n")

start_time_detect_language = time.time()
language_info = model.detect_language_multi_segment(
    audio_file,
)
print(language_info)
print("\nDectech language time:", time.time() - start_time_detect_language, "\n")

batched_model = BatchedInferencePipeline(
    model=model,
    vad_device="cuda",
    use_vad_model=True,
    vad_onset=0.1,
    vad_offset=0.1,
    language=language_info.get("language_code"),
)

segments, info = batched_model.transcribe(
    audio=audio_file,
    without_timestamps=without_timestamps,
    word_timestamps=word_timestamps,
    batch_size=batch_size,
    beam_size=beam_size,
    patience=patience,
    task=task,
    language=language_info.get("language_code"),
)

segments = list(segments)

print("info: ", info.language, info.language_probability)
print("Processing time:", time.time() - start_time, "\n")

for segment in segments:
    print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))

[drscotthawley]

In running this, I get the error:

Traceback (most recent call last):
  File "/Users/myusername/test_batched.py", line 33, in <module>
    language_info = model.detect_language_multi_segment(
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'WhisperModel' object has no attribute 'detect_language_multi_segment'

Commenting out all relevant parts that relate to that error, we find another error:

Traceback (most recent call last):
  File "/Users/myusername/test_batched.py", line 39, in <module>
    batched_model = BatchedInferencePipeline(
                    ^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: BatchedInferencePipeline.__init__() got an unexpected keyword argument 'vad_device'

[MahmoudAshraf97]

please use the latest version, you are using a non-updated version