Duplicate sentences with large-v3

[NingLu]

Hi,

I see duplicate sentences when using large-v3, the issue can be reproduced no matter I set condition_on_previous_text to True or False, any suggestion on how to fix it?

Here's the codes

def transcribe_audio(audio_file, beam_size=5, vad_filter=True):
    """Transcribe audio using Faster-Whisper"""
    print(f"Transcribing: {audio_file}")
    
    start_time = time.time()  # 新增：记录开始时间
    segments, info = whisper_model.transcribe(
        audio_file,
        beam_size=5,
        language=LANGUAGE,
        vad_filter=True,
        word_timestamps=True,
        condition_on_previous_text=False,
        vad_parameters=dict(min_silence_duration_ms=200),
    )
    
    segments = list(segments)
    transcribe_time = time.time() - start_time  # 新增：计算耗时
    
    print(f"Language: {info.language} (probability: {info.language_probability:.2f})")
    print(f"Segments: {len(segments)}")
    print(f"Transcription time: {transcribe_time:.2f}s")  # 新增：打印耗时
    
    return segments, info, transcribe_time  # 修改：返回值增加 transcribe_time

The segments output

TRANSCRIPTION WITH SPEAKER DIARIZATION
================================================================================

[18.62s -> 22.83s] SPEAKER_2
  莫子缓缓地说道

[23.52s -> 25.68s] SPEAKER_4
  恒济原将继续

[26.90s -> 29.32s] SPEAKER_4
  宇宙是一台机器

[29.32s -> 33.40s] SPEAKER_4
  我造出了这台机器

[33.98s -> 36.84s] SPEAKER_4
  恒纪元将继续

[38.22s -> 40.82s] SPEAKER_4
  宇宙是一台机器

[40.82s -> 45.02s] SPEAKER_4
  我造出了这台机器

It is not caused by speaker diarization as I print the segments before invoking speaker diarization