Title: Inconsistent transcription of quiet and louder speech
Title: Inconsistent transcription of quiet and louder speech, with long sections being skipped
The overall transcription quality seems quite good, but the model appears to be highly inconsistent.
In general, quiet speech is often not recognized. However, in some cases, the model successfully transcribes speech even when the volume is very low.
The more confusing issue is that, when transcribing recordings made in the same environment and under similar conditions, the model may successfully recognize quiet speech but fail to transcribe other speech that is noticeably louder than the quiet speech it recognized.
When the model fails to recognize the louder speech, it often does not skip only a short sentence or a few seconds. In many cases, it fails to transcribe a long continuous section of the recording, even though clearly audible speech is present throughout that section.
This happens quite frequently, so the transcription result does not seem to depend solely on the volume of the speech. The behavior appears unstable and difficult to predict.
Has anyone else experienced this issue? Are there any recommended settings, decoding parameters, or audio preprocessing methods that can improve transcription consistency and prevent long sections of speech from being skipped?
Thank you very much for your valuable feedback. If possible, could you please share the audio file you used for testing? We would like to investigate this issue further and work on improving the model's performance.
Thank you for your response.
Unfortunately, I cannot share the recording because it contains confidential internal company meeting discussions. I understand that having the original audio would be helpful for investigating the issue, but I hope you can understand that I am unable to provide it due to privacy and confidentiality concerns.
Thank you for your understanding.