Checking a Cloned Voice Word by Word

The narrator of one of my Italian channels is a clone of a real voice, made with the open-source VoxCPM2 model running on my own graphics card. Most of the time it sounds good. Now and then it opens a sentence with a syllable that belongs to no language, swallows the first word, or garbles one in the middle. A viewer notices that within seconds, so every sentence has to pass a check before it goes into a video. The script is cut into sentences and each sentence is generated as its own take. Two Whisper speech recognizers, the small and the medium model, transcribe the take, and I compare what they heard with the script, word by word. A take is rejected for a missing word, an extra word, or a word the recognizer isn't sure about. A sentence that is 93 percent right can still contain the one word that sounds wrong, so close doesn't count. Then the take is trimmed, and the trimmed audio goes through both recognizers a second time. It's accepted only if both return exactly the script. Two details came from real failures. Recognizers ignore short sounds that aren't words, while a listener hears every one of them. So I measure the audio level in the pauses between sentences and cut any isolated burst. The other detail is numbers: a script that says "9 minutes" comes back from the recognizer with the digit 9, the voice says "nove", and the comparison fails. I convert digits to words on both sides before comparing. Sometimes the model drops the same word on every attempt. One sentence began with "Immergiti" and lost it in all six takes, so I replaced the verb with "Scendi" and the next take passed. After eight failed attempts a sentence is split into two shorter ones and tried again. This makes production slower. A short video with five sentences can need a dozen takes or more. I accept that, because the voice is the first thing a viewer judges and computer time is cheap. The step I'd copy first is the second pass on the trimmed audio. Trimming can introduce its own errors, and the trimmed file is the one that gets published.