Making Reliable Word Recordings with AI Voice Models

Summary

The author describes building a spelling game that needed one clear WAV recording for every word, then working through the problems created by AI-generated speech. Recordings made with audio.cpp were produced in English, French, and Norwegian; the workflow moved from OmniVoice to VoxCPM2 for voice creation because of licensing requirements, and used a VoxCPM2 reference recording with Chatterbox to generate sentences. A carefully written teacher prompt and a reference sentence were used to encourage calm, deliberate articulation, with the reference later slowed to 85% of its original tempo. Generating isolated words proved unreliable: some outputs lost the beginning, ended too early, or included part of the preceding word. Adding a sentence template gave the model more context, but extracting the repeated final word remained difficult. Silence detection could identify quiet intervals, but could not determine whether a pause separated words or occurred inside a word. Rules based on the last or longest pause, remaining duration, letter count, or volume thresholds each failed on different examples. The author therefore built audio-lab, a browser tool for waveform inspection, zooming, slowed playback, manual selection, silence overlays, and loading machine-generated selection JSON. Estimating where the introductory sentence should end and searching nearby pauses improved the heuristics, but the author emphasizes that this was rule refinement rather than model training. Changing the sentence wording and punctuation also improved generation; in English, a template beginning with “Please say the word” worked best in the reported tests. A separate cleanup step removed some unwanted lead-in sounds, while preview WAV files and JSON made the result inspectable. Whisper and nb-whisper supplied useful mismatch signals for batch review, although transcription flags still required listening: seven Norwegian flags produced only one judged error, while seven English flags were accepted. Licensing the model weights remained important, with OmniVoice Word-Control marked CC-BY-NC-4.0 and the cited Chatterbox and VoxCPM2 model cards listing MIT and Apache-2.0 licenses. The article concludes that AI accelerated the creation of small diagnostic tools, but human listening and judgment remained necessary.