Audio brief

Two jobs, and the second one is the hard one. Rendering text to speech is a command. Writing something worth listening to is the craft, and skipping it produces a file that is technically audio and practically unlistenable.

Render it

brief-audio script.txt                  # -> script.mp3
brief-audio -v am_michael notes.txt     # different voice
echo "some text" | brief-audio -o out.mp3

Kokoro-82M, Apache 2.0, local, no API key and no network after the first run caches the model. First run installs espeak-ng, ffmpeg and a Python 3.12 venv at ~/.chewbacca/audio-venv, which takes about a minute. Every run after is faster than real time.

Voices: af_heart (default, highest quality grade), af_bella, af_nicole, af_sarah, am_michael, am_puck, am_adam, bf_emma, bm_george.

Do not reach for a paid TTS API. That was the first instinct on 2026-09-20 and it was wrong: it needed a key Caleb did not have, so the work stalled on a credential for something a local Apache-2.0 model does well. If a hosted voice is ever specifically wanted, that is a deliberate choice, not the default.

Write it for the ear first

The source analysis is written for eyes that can scan, re-read and skip. None of that exists in audio. A listener gets what you give them, in the order you give it, exactly once. Rewrite before rendering. Never pipe a document straight in.

Front-load the conclusion. No warm-up. The first sentence is the finding.

Kill everything visual. No tables, no bullets, no headers, no URLs, no markdown, no file paths. A table read aloud is noise. If a table mattered, say the two numbers that carried it.

Speak the numbers. 81,460 becomes "eighty one thousand, four hundred and sixty". 18.5% becomes "eighteen and a half percent", or better, "roughly one in five". $3K/mo becomes "three thousand a month". Round hard: precision the listener cannot write down is wasted.

Say the headline number twice. Once in passing, then "let me say that again" and once slowly. They cannot scroll back.

Respell anything the voice will mangle. a16z becomes "a sixteen z", CPG becomes "C.P.G." with the periods in, and NaN becomes "N-A-N". Check acronyms, tickers, domain names and anything with digits jammed into letters.

Signpost every transition. "First thing." "Second thing, and this is the interesting one." "Okay, now the dataset." These are filler on the page and load bearing in audio, because they are the only structure a listener gets.

Short sentences, because there is no re-reading. A clause that needs a second pass is a clause that is lost.

Say who is who. "Sam texted you" not "he said". Pronouns drift fast when nothing is on screen to anchor them.

Length follows the activity. Laundry or a walk is four to six minutes, which is about six hundred to nine hundred words at a hundred and fifty words a minute. A commute takes more. Check the word count before rendering, not after.

End on the action, not a recap. They just heard it.

The loudness step is not optional

brief-audio normalizes to -16 LUFS automatically. That is the podcast standard and it exists because raw Kokoro output sits near -26 dB mean, which is audible in headphones in a quiet room and completely gone next to a washing machine or road noise. If you ever render outside this tool, run the same pass:

ffmpeg -y -i raw.wav -af "loudnorm=I=-16:TP=-1.5:LRA=11" -codec:a libmp3lame -b:a 128k out.mp3

Deliver it where they will actually play it

Put the file somewhere reachable without a file browser, usually ~/Desktop or ~/Downloads, and say the filename and the runtime in the reply. Keep the script next to the audio so it can be corrected and re-rendered without rewriting.

Gotchas that cost real time

• Python must be 3.12. torch has no wheels for 3.13 or 3.14, and the system Python on this machine is 3.14.
• uv pip install --python wants the venv directory, not <venv>/bin/python. That path is a symlink to the uv-managed interpreter, so uv resolves it to the real binary and then reports "No virtual environment found".
• kokoro shells out to uv pip install for a spaCy model on first use. It reads VIRTUAL_ENV, so rendering from a directory with no .venv fails unless that variable is exported. brief-audio exports it and pre-installs the model.