Complete silence is always hallucinated as "ترجمة نانسي قنقر" in Arabic (github.com)

Article summary

The article discusses how the Whisper AI model hallucinates a specific Arabic phrase, "ترجمة نانسي قنقر" (Translation by Nancy Qunqar), when given complete silence as input. This issue is not unique to Arabic, as similar hallucinations occur in other languages, such as German and Norwegian. The problem is attributed to the model's training data, which includes unofficial subtitles from movies that often have a "translated by" string at the end. The issue has been reported since at least February 2024 and affects other AI models, including ChatGPT's voice mode.

Main themes

• AI model hallucinations
• Training data quality
• Copyright laws
• AI industry ethics
• Language processing biases
• Machine learning errors

What commenters say

• The Whisper AI model's hallucination of a specific phrase when given silence is due to its training data, which includes unofficial subtitles from movies.
• The issue is not unique to Arabic and occurs in other languages, such as German and Norwegian, where the model hallucinates different phrases.
• The use of unofficial subtitles from movies as training data may be a violation of copyright laws.
• Some argue that the AI industry's use of copyrighted material without permission or compensation is a form of piracy.
• Others believe that the AI industry's actions are justified, as they are creating new and innovative products.
• The issue highlights the need for more careful curation of training data to avoid biases and errors.
• The problem may be solved by using techniques such as voice activity detection (VAD) to remove silence from audio files.
• The AI industry's actions may have significant implications for copyright laws and the future of creative works.