When recognizing Chinese, if the input audio is silent or consists only of background noise, the model often outputs the following unrelated fixed text

[ojama]

Environment
• Model: whisper-large-v3

Bug Description

When recognizing Chinese, if the input audio is silent or consists only of background noise, the model often outputs the following unrelated fixed text:

“请不吝点赞 订阅 转发 打赏支持明镜与点点栏目”

This output is completely unrelated to the input audio.

Expected Behavior

For silent audio or background noise input, the model should:
• Return an empty string ("") or a special token (e.g. [NO_SPEECH]),
• Or explicitly indicate that no human speech was detected.

It should not output irrelevant text such as advertisements or unrelated phrases.

To Reproduce
1. Load whisper-large-v3
2. Provide an input audio clip that is either silent or only contains background noise
3. Observe that the model outputs the unrelated fixed phrase

Possible Cause
• The training dataset may contain a large number of samples with background music + subtitles, where the subtitles do not match the actual audio content.
• Lack of volume-based filtering or VAD (Voice Activity Detection) during training, causing non-speech segments to be paired with subtitle text.
• Because such noisy data likely appeared frequently, the model learned to output these fixed high-frequency phrases in non-speech conditions.

Suggestions for Improvement
• Apply volume-based filtering or VAD to clean training data
• Assign no_speech labels for non-speech segments instead of pairing them with subtitle text
• Explicitly label audio source types (human speech / background music / environmental noise) in the training data so the model can distinguish them properly