How the audio logs were made

There are six audio recordings in Chapter One. All six are generated on the server from the same transcript the game shows you, which is the single most important decision in the whole pipeline: the recording cannot drift from the text, because the text is the input.

Voices

Each character is synthesised with a different neural voice, at a different speaking rate:

• Kessler, a tired physician, slightly slower than natural.
• Aldaine, thirty-one and awake at four in the morning, at natural pace.
• Palliser, a records clerk with nine years of practice at being clear.
• Reeve, a night security officer from the north of England.
• Vance, the director, unhurried, never louder than she needs to be.

Stage directions

Transcripts contain bracketed directions — a pause, a short laugh, a telephone ringing four times below, footsteps and a chair and a door. Every one of those has an explicit rule in the build script. An unknown direction is a hard error that stops the build, rather than a guessed one-and-a-half-second gap, because a guessed gap is how a recording quietly stops matching its own transcript.

Some directions are pinned to a clock time. Reeve's log describes a telephone ringing at 1:12. The build lays out the whole timeline, works out where the preceding speech actually lands, and pads the silence so the ring falls at seventy-two seconds exactly. If the speech has already overrun that mark the build refuses to run, which is how it should behave — the alternative is a description that lies about its own audio.

Room noise

Every recording has a documented background: rain on a window and a radiator ticking, a ventilation fan and a pen, a typewriter carriage returning once, tape hiss and the whine of a monitor, a clock and one car far off, mains hum and nothing else at all. Each of those is synthesised for that recording specifically and then gain-staged to an absolute level, so the room sits between fifteen and nineteen decibels under the voice in all six. That number is measured on the finished file, not estimated.

The tape

The mix then goes through one chain, once: a bandwidth of roughly 95 Hz to 5.4 kHz, the automatic gain a dictation machine applies whether you want it or not, a little wow, a small room, and a limiter. The degradation is deliberate and identical across all six, so the recordings sound like they came off the same machines in the same building.

What went wrong

The first pass came out at a peak of −11.8 dBFS, which is far too quiet, because the compressor was squashing the speech and nothing was making up the gain afterwards. The fix was to measure and then apply, rather than to guess a boost.

The second pass buried the dialogue under the rain. Measuring per-second levels across the whole file showed the room sitting three decibels under the speech instead of the intended sixteen. The third pass overcorrected and put the room twenty-six decibels down, which is inaudible. The number in the script now is the one that measured correctly.

A subtler one: the tape chain originally rolled off below 165 Hz, which silently deleted the mains hum that two of the recordings are largely made of. The ward recording is supposed to be a room with a hum in it. Lowering the corner to 95 Hz gave it back.

Every recording ships its transcript

This is a rule rather than a nicety. No clue in this game exists only in sound. Every recording has a full written transcript on the same page, every surveillance clip has a written scene description, and the puzzles that involve audio have a route through them that does not.

Reeve's recording is five and a half minutes long, and three minutes and nine seconds of that is tape hiss after he leaves the room. That is what the transcript says happens, so that is what the file does.