PärPod Autocast
PärPod Autocast
PärPod Autocast
The Stack Report: A Voice Model That Fits In A Pocket, And FastAPI Takes Its Checkbox Back
8m · Oct 28, 2026
The Stack Report: A Voice Model That Fits In A Pocket, And FastAPI Takes Its Checkbox Back

The Stack Report: A Voice Model That Fits In A Pocket, And FastAPI Takes Its Checkbox Back

A cloned voice, on a laptop CPU

Kyutai's Pocket TTS now ships inside Transformers 5.21. It is a 100 million parameter text-to-speech model that runs faster than real time on an ordinary CPU, with no GPU and no MLX port needed. It can also clone a voice from a short recording.

Today is Wednesday the 28th of October, seven days on from the previous episode.

The connection is the x feed. It is fully local and ephemeral, and it reads in Kokoro's bm_lewis voice, the one whose mispronunciations you are already patching with a respelling lexicon. Pocket TTS is a second local engine with a different architecture. It generates continuous audio latents at 12.5 frames a second and decodes them to 24 kilohertz through its own codec, with a separate head that decides when the utterance ends. Because it runs on the CPU, it would not compete with ordagrant or a VLM for Metal memory. That matters on the one Mac where wired memory has frozen everything before.

The cloning is the more interesting half. A local engine that speaks in a voice you choose, from a recording you already have, would make a private feed sound like something other than a stock narrator. It would cost nothing per minute and need no subscription.

Three things the notes do not tell me:

Your own rule settles where this starts: voicelab first, with Kokoro and the current Inworld voice side by side on a quickcheck page, and one passage of at least ten minutes. The weights wait for your yes. That is an afternoon, and most of it is listening.

The same release adds Gradium Phonon TTS, a sibling built on the same recipe. It cannot clone from a recording, so it is a footnote.

A second OCR candidate for parocr

Transformers 5.21 also adds MiniCPM-OCR from OpenBMB. It uses the MiniCPM-V 4.6 architecture, with a SigLIP vision encoder in front of a Qwen 3.5 language model, and it is tuned for text-dense images: scanned pages, receipts, tables and screenshots. Parocr's compare mode already has UnlimitedOcr parked as a candidate since last week. This would be the second, and the more interesting one, because the MiniCPM-V family has a history of turning up in MLX-VLM soon after release, and parocr's own engines are MLX.

Your scans eval is still waiting on a default ruling between tesseract and Apple Vision, so put this into the same seeded rig, not into production. If it beats Apple Vision on the T036 pages, that ruling changes. That is a session once an MLX version exists, and an afternoon of plumbing if you want it before then.

What happened to last week's leads

Most of them landed in todo files, which is where leads belong:

On Imagebook I owe you a correction of a different kind. For two weeks running I led with a way to search your photos by image, not by caption. On the 23rd Imagebook was archived. A scout who keeps pitching a door on a house that is being emptied should have checked the house first.

Heads up

FastAPI took back the checkbox. Last week I told you an omitted checkbox now reads as false. Version 0.162 reverses that. An omitted true-or-false form field now uses its declared default, and if it has no default, it is a missing-field error. In practice, an HTML checkbox declared without a default of false now rejects every form submitted with the box unticked.

Version 0.160 adds a second change. A query, header, cookie or form model that forbids extra fields now rejects sibling parameters it does not declare, even ones another parameter declares. Both of these hit forms, so Deadline's booking pages and PartyPär's admin again. The advice is the same as last week, only firmer: when a lockfile refresh carries FastAPI past 0.159, a form test has to submit an unticked box.

Anthropic SDK 2.0. On the 21st Anthropic released API version 2026-10-21 and a Python SDK 2.0.0 that sends it by default. Both have breaking changes that do not depend on the API version. Ask, Capture's categorizer, the Builder and parpod-core all call Claude through that SDK. Director's exp 096 rig already pins below 2, which is the right instinct. The other services only stay on 1.x if their locks say so, and I have not checked that they do. The trap is the same one as last week's certifi sweep: a blanket upgrade pulls a major version along with it. Your sweep refreshed certifi alone in about thirty repos. That was correct, so keep doing it that way.

The certificate story is not finished. uv 0.13.5 moved the managed Pythons to builds whose bundled pip trusts the corrected root list, and Python shipped security patches on the 21st and again on the 25th. Two legs remain:

Popcorn2's base images are a separate question, and floor already has it open.

Pocket ID 2.19. Popcorn2 runs 2.18.0. Version 2.19 rejects cross-origin state-changing requests, closes an email enumeration path, and moves new passkeys to WebAuthn's recommended algorithms. 2.19.1 rebuilds it on a newer Go and base image. It is your login plane, so take it in one sitting per the image-mirror rule, then do the passkey walk again.

The rest, briefly

If you look at one thing

Put Pocket TTS next to bm_lewis in voicelab and listen for ten minutes. If it holds up, the x feed gets a voice of your own and still never leaves the Mac.

PärPod