Read the explanation
The source divides its normalized utterance into an aspiration region before zero point two two and a voiced region afterward. Inside the voiced region it sets g equal to t minus zero point two two, divided by zero point seven eight. The formant interpolation weight is one divided by one plus exponential of negative eight times g minus one half. At g zero the weight is about zero point zero one eight, not zero. Move through the voiced interval from g zero to one. The marker follows the exact source sigmoid. It passes through one half at the interval midpoint and approaches about zero point nine eight two at the end. The horizontal coordinate is normalized voiced progress. Video pacing follows narration rather than claiming the same duration as the synthesized audio. The weight never reaches exact zero or one on this finite interval. Therefore the calculated formants start and finish slightly inside the selected endpoint frequencies. The page rounds the returned frequency values afterward. Our graph deliberately shows the unrounded interpolation expression, not a quantized spectrogram or a measured pronunciation. Select first formant endpoints eight hundred and three hundred hertz, and second formant endpoints twelve hundred and twenty four hundred hertz. These are two independent source control pairs. Both use the same sigmoid weight. At the beginning of the voiced interval, the first formant is about seven hundred ninety one hertz and the second about twelve hundred twenty two hertz before rounding. Advance the normalized voiced progress from zero to one. First formant equals eight hundred minus five hundred times the weight. Second formant equals twelve hundred plus twelve hundred times the weight. The two colored markers move on the same hertz scale: one decreases and the other increases. This is the source frequency schedule before integer rounding, not an audio-spectrum measurement. At the end, first formant is about three hundred nine hertz and second about twenty three hundred seventy eight hertz. Neither exactly reaches its selected target because the sigmoid endpoints are finite. The page uses illustrative source-filter controls and fixed phoneme segment labels, so these schedules do not prove a faithful spoken word or a measured human vocal tract. Pitch uses a different expression. Set base pitch to one hundred fifty hertz and contour to thirty hertz. During the voiced region, the source pitch is base plus contour times normalized progress. Initially it is one hundred fifty hertz. Before the voiced region, the source instead sets pitch to zero and labels aspiration; this scene starts at the voiced boundary. Increase normalized voiced progress from zero to one. Pitch rises linearly from one hundred fifty to one hundred eighty hertz. This line does not use the sigmoid formant weight. The graph separates the pitch contour rule from the two formant schedules. Returned pitch is rounded to an integer afterward, so the smooth line illustrates the underlying formula rather than every rounded audio control value. The final underlying pitch is one hundred eighty hertz. Formants and pitch play different roles in the source model, even though their diagrams all use frequency units. Preset labels and export targets remain fixed in this tool, including presets named for other utterances. The video explains actual control arithmetic and preserves the original Web Audio, WAV and JSON workflows without claiming calibrated speech synthesis.