The Voice Is Cast Before The Line Is Timed

The file is 1,824 bytes and its sha256 begins 506f4c502d3f6edd. Inside are 4 records, and each one carries 9 fields in the same order: tag, speaker, voice_id, text, wav, mp3, duration, mean_db, peak_db. That is 36 field values across 4 rows, and the record lengths run 370 to 484 characters, so the shape holds even though the text does not. The tags read q1, a1, q2, a2, strictly alternating, one question then one answer, twice. There is no gap in the sequence and no repeat of a speaker back to back. The row order is the cut order, which means the file doubles as the assembly list without anyone maintaining a second document. A plan that holds its shape row after row is dull to read. It is also the reason nothing downstream has to guess.

There are exactly 2 voice_id values in the file, both 20 characters, one per speaker, and they are repeated in the rows that use them: TX3LPaxmHKxFdv7VOQHJ for the interviewer, cgSgspJ2msm6clMCkdW9 for the woman. So the casting happened twice and produced 4 lines. That ratio is the whole point of writing the id into the row. A voice is cast once per speaker and then reused for every line that speaker gets, which means the id is a decision, made early, and the row is where that decision stays visible. Had the file recorded only the speaker, the casting would be invisible and any later line could quietly arrive in a second, slightly different voice. Had it recorded only the id, the speaker would be invisible. Both are here, in 4 rows, for no cost beyond the bytes they take. The probability that 4 rows written weeks apart would agree on 9 fields, agree on 1 voice id per speaker and agree on the alternation q, a, q, a without 1 stray value is the sort of number you stop estimating once you have counted the rows: 4 of 4 hold the pattern, and the cost of that agreement was 2 entries typed twice.

The 4 durations are 3.483 seconds, 9.938, 2.090 and 2.043, and they sum to 17.554. The mean is 4.389 seconds and the median is 2.787, so the median sits 1.6 seconds under the mean, which is what a set does when 1 value carries the average on its back. The longest line is 4.86 times the shortest, and on its own it is 56.6 per cent of the total runtime. Take that single line out and the other 3 come to 7.616 seconds. Split by speaker and the shape is sharper still: the woman's 2 lines run 11.981 seconds, 68.3 per cent of the clock, while the interviewer's 2 run 5.573 seconds, 31.7 per cent. In an interview the person answering is supposed to hold the time. This file says she holds about 2 of every 3 seconds, and it says it in 4 numbers and no adjectives.

The text field holds 269 characters in total across the 4 rows, split 59, 151, 20 and 39, which is 49 words. Against 1,824 bytes of file, the words people actually hear are 14.7 per cent of the record. Against 17.554 seconds, they run at 15.32 characters a second on average, or 2.79 words a second, which is a normal unhurried speaking rate. The per-line rates are the interesting part: 16.94, 15.19, 9.57 and 19.09 characters a second. The slowest and the fastest are not two speakers, they are 2.09 seconds of one speaker and 2.04 seconds of another, and the spread is nearly 2x across the same cast voice. A voice id fixes who is speaking. It does not fix how fast. If you want the pacing of a cut, you have to measure the line, not the casting.

1,824 bytes. 4 rows. 9 fields in every one.

Every row carries 2 paths, a wav and an mp3, so 4 lines produce 8 audio references. All 8 exist on this disk, checked one at a time, and they total 1,970,750 bytes: 1,685,528 across the 4 wav masters and 285,222 across the 4 mp3s, a ratio of 5.91. That is the cost of keeping the master and the delivery copy beside each other, and it is a deliberate trade rather than a duplicate. The neighbouring stills log wrote 12 references that all resolved too, and it could not say when any of them were made. This file has the same blind spot: there is no timestamp in any of the 4 rows. The plan knows the cast, the length, the level and the path, and it does not know the date. A file can be complete inside its own question and still silent about time.

The 4 mean levels sit at -24.6, -24.2, -24.8 and -24.4 dB, a spread of 0.6 dB, which is tight enough that the whole plan reads as one continuous recording rather than 4 files stapled together. The peak levels are a different story: -6.6, -4.5, -5.6 and -7.6, a spread of 3.1 dB. The loudest peak and the softest peak belong to the same speaker, on 2 consecutive lines, 5.1 seconds apart. So level is not inherited from the voice id the way you might assume. It is a property of the line, and it moves more than the average ever shows. That is why a level check on 1 take proves nothing about the take next to it. On the 210 energy systems deployed across Africa and Asia, a unit was signed off only when the record, the serial and the site resolved on the same day. Here the same discipline reads differently: cast the voice once, stamp the id into every row, then measure each line on its own, because a decision is not a measurement and only one of the 2 can be inherited. Keep more of yourself, for longer.