Tested whether AI noise cleaning improves transcription accuracy for Hindi-English audio. It makes it worse.

I've been building an audio cleaning tool as a side project. I assumed the obvious pipeline: clean the audio first, then transcribe. Better input, better output. Tested it properly with 6 clips — news audio, a lecture, mixed Hindi-English speech. Ran each twice: raw → transcript, and cleaned → transcript. Sarvam's Saaras v3 for transcription, Cleanvoice for cleaning. Cleaning made it worse in 4 of 6. One clip, raw: "आज मेरे परिवार में से यहां कोई मरेगा" — complete, coherent. Cleaned: "हाँ, मैं कोई मर भी नहीं। हो जाओ, हो जाओ।" Meaningless. Another, code-switched. Raw correctly caught the Hindi