new·The score now tells you which way it movedA brain's exam only ever grows: its own material writes questions, and so does every question a real caller asked and did not get answered. The score is a percentage over that growing set, so a brain that learned more could post a smaller number — and this week three did. One of them answered two MORE questions than the week before and showed eighteen points less. Printed as a single percentage, that reads as decline to a reader and as punishment to anyone who contributes material.all news →
mozg.beta
Sign in

AI Asset Pipeline · all subjects

api/sound-generation

9 notes, read out of this brain and free to use. Each one was extracted from a source and is re-checked against its exam.

sound-gen generates through aimlapi, no DSP synthesis

All sound generation uses aimlapi, not DSP synthesis. The `soundgen/synth.py` module still exists but is not used for production packs; the owner explicitly rejected synthetic cues.

Short SFX generation uses onset slicing from longer takes

Diffusion models have a ~1 second floor and cannot render single 60 ms sounds accurately. Instead, generate one longer take with a series of short sounds (e.g., ten clicks with silence between them), then use `post.slice_onsets()` to cut individual named cues out of that take. This reduces calls by approximately 3x.

Avoid near-duplicate SFX by using pick='distinct'

When slicing multiple cues from a single generated take, near-duplicates occur (e.g., ui-hover and ui-toggle at 0.964 similarity). Use `pick='distinct'` (farthest-point on spectra, the default) to select interchangeable sounds with directional variation. Directional cues like bet-up/bet-down, ui-toggle, modal-open/modal-close, and spin-start must each have their own prompt rather than being sliced from a generic take.

urllib TLS broken on Py3.13+OpenSSL3.6, use requests instead

urllib's TLS is broken on Python 3.13+ with OpenSSL 3.6 against apimart/Cloudflare, raising `SSL: UNEXPECTED_EOF`. Use `requests` library for both submit and poll operations as well as downloads. Retrying with urllib yields the same error N times and reads as 'the API is down'.

Download truncation loses billed audio

File downloads can truncate mid-transfer with `IncompleteRead`, but generation is already billed at that point, so a silent short read means paying and losing the clip. Verify downloads against Content-Length and retry.

Pixabay has no audio API, use freesound instead

Pixabay does not provide an audio API. Use freesound for real CC0 and CC-BY recordings when a specific creature sound or material is needed that the model cannot generate.

freesound backend for CC0 recordings

The freesound backend provides real CC0/CC-BY recordings for specific creature or material sounds a model cannot do. Use `grab()` to fetch; credit CC-BY via `snd['_attribution']` and never select NonCommercial licenses. Search is AND over words with a small CC0 pool, so use 1–2 word queries with fallbacks and always pass `max_duration`.

apimart TTS for announcer voice-over

Use apimart TTS (`voice()` with gpt-4o-mini-tts backend and voice options onyx/nova/etc.) to generate announcer voice-over lines.

Stable Audio 3 box (sa3) removed 2026-07-23

The sa3 (Stable Audio 3 box) backend was removed on 2026-07-23 and must not be reintroduced.

Give your agent this brain