Batched Faster Whisper misses few sentences in the audios

[toanhn]

I'm using Batched Faster Whisper to transcribe the audio files. But it sometimes misses some sentences in the audio files. Is there any argument to control this behavior, I try "vad_onset" and "vad_offset" arguments but it works occasionally. Do you know any way to resolve this issue? Thank you so much.

[x86Gr]

Can you paste the code you're using?

[toanhuynhnguyen]

from faster_whisper import BatchedInferencePipeline, WhisperModel

model_size = "large-v3"
device = "cuda"
compute_type = "float16"

batch_size = 16
beam_size = 3
without_timestamps = True
word_timestamps = True
task = "transcribe"
language = "en"
initial_prompt = "Microsoft, OpenAI"  # some terminologies, proper noun...

audio_file = "test.wav"

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

segments, info = batched_model.transcribe(
    audio=audio_file,
    batch_size=batch_size,
    beam_size=beam_size,
    without_timestamps=without_timestamps,
    word_timestamps=word_timestamps,
    task=task,
    language=language,
    initial_prompt=initial_prompt,
)

segments = list(segments)

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

@x86Gr When I remove initial_prompt then it can capture almost all sentences in the audio, looks like initial_prompt causes missing some sentences in the audio. But I need to capture some terminologies, and proper nouns in the audio.