Result Segment with wrong duration of audio file.

[tcl8273]

I am using

uncoverV2.mp4

whisper to get segments but I see the end time of the final segment is the wrong duration

{'text': " Nobody sees, nobody knows We are a secret, can't be exposed That's how it is, that's how it goes", 'segments': [{'id': 0, 'seek': 0, 'start': 0.0, 'end': 4.5600000000000005, 'text': ' Nobody sees, nobody knows', 'tokens': [50364, 9297, 8194, 11, 5079, 3255, 50592, 50592, 492, 366, 257, 4054, 11, 393, 380, 312, 9495, 50864, 50864, 663, 311, 577, 309, 307, 11, 300, 311, 577, 309, 1709, 51140], 'temperature': 0.0, 'avg_logprob': -0.39771583676338196, 'compression_ratio': 1.1428571428571428, 'no_speech_prob': 0.1816209852695465}, {'id': 1, 'seek': 0, 'start': 4.5600000000000005, 'end': 10.0, 'text': " We are a secret, can't be exposed", 'tokens': [50364, 9297, 8194, 11, 5079, 3255, 50592, 50592, 492, 366, 257, 4054, 11, 393, 380, 312, 9495, 50864, 50864, 663, 311, 577, 309, 307, 11, 300, 311, 577, 309, 1709, 51140], 'temperature': 0.0, 'avg_logprob': -0.39771583676338196, 'compression_ratio': 1.1428571428571428, 'no_speech_prob': 0.1816209852695465}, {'id': 2, 'seek': 1000, 'start': 10.0, 'end': 31.0, 'text': " That's how it is, that's how it goes", 'tokens': [50364, 663, 311, 577, 309, 307, 11, 300, 311, 577, 309, 1709, 51414], 'temperature': 0.0, 'avg_logprob': -0.247916613306318, 'compression_ratio': 1.0909090909090908, 'no_speech_prob': 0.004447056911885738}], 'language': 'en'}

The audio durations are 15s

[jianfch]

Empirically, I found that this tends to go away with the large model. You can also just add some lines to clamp the max timestamp to the duration of the audio. Another way is to suppress any timestamp tokens that is greater than the audio duration at the decoding stage.

[tcl8273]

I used a large model to get transcribe.

[tcl8273]

when I used Ubuntu I see the different duration (24s) instead of 15s:

Could you help me resolve this error?
Thank you.

[jianfch]

You can implement the latter options I mentioned. or just use stable-ts

[r4gor]

I am having the same problem