The Stack Report: The Voice Plan You Already Pay For Can Now Hold A Conversation

Two voices, one breath

Ten days ago I pointed you at a research model that generates a conversation between several voices as one continuous act. The advice was one render to find out whether the April problems were fixed. Since then, a vendor you already pay every month has shipped the same idea as a finished product.

It is Wednesday, 30 September, and 10 days have passed since the last one.

ElevenLabs released Eleven v4 on Monday. There are two variants. The full model runs through their Text to Dialogue interface and is built for long-form and multi-speaker audio. The Turbo variant is aimed at real-time agents, and that one is not for you. The full model interests me for two reasons, and neither is the headline quality claim.

First, dialogue requests can now carry the text before and after a passage, plus the IDs of earlier and later requests, specifically so that consecutive generations sound continuous. Your pipeline renders episode by episode and segment by segment, and every segment boundary is a place where prosody resets. This is the first time any engine you use has offered a mechanism aimed at that seam.

Second, the economics have not changed since I last raised them. Your ElevenLabs credits expire monthly, and under-use is the failure mode you have on record for that plan. Today those credits mostly buy stings and sound effects. A two-voice sample on v4 costs you nothing at the margin.

Two cautions. v4 also claims better voice cloning across more than 90 languages. Swedish is the lane where your Inworld clone went wrong, so it deserves a separate test, but a claim is still just a claim. And the rule has not moved: nobody, me included, decides how this sounds without you listening to it. One dialogue render against the same passage on Inworld, both labelled and sent to the temp feed. That is an afternoon, and most of it is listening.

FastAPI now ships its own telemetry

FastAPI 0.142.0 added native OpenTelemetry support. The release note is a single line with no details, so I will tell you what it means for you and not what the feature contains.

This week the telemetry-everywhere lane was formally closed. The utilities repo merged the canonical telemetry setup, parkit merged the shared version, and between them seven decisions are waiting for you, including two about the telemetry template itself. Every FastAPI service you own carries a copy of that setup file. Some of that copy exists only because the framework offered nothing of its own.

The question is how much of the template the framework can now own. If native support covers request spans cleanly, the per-project copy gets smaller and there is less to drift. The part I would not expect it to handle is your popcorn2 exporter path, which sends over a Unix socket and not a network port. Answering this is one reading session before you rule on the template, and the timing is right: you should not decide a template's shape in the same week the framework moved underneath it.

The item left on the imagelab list

Last time I said the MLX vision library now covered most of what you were pricing on 11 September. One item was missing: human-specific pose and body-part segmentation. Version 0.7.2 added a port of Meta's Sapiens2, a family of dense prediction models built for people, covering pose, body-part segmentation, depth and surface normals. The same release added SAM 3D Objects.

That completes the list, on Apple Silicon, in the runtime you prefer. The imagelab question from last time, whether that lane still needs a PyTorch checkpoint, now has one fewer excuse to go unanswered.

The same cycle also added Qwen Image 2.1 locally, for both generation and editing. One line on that, because the connection is tempting and wrong. Framkallaren is replacing a one-click enhance tool for Årebladet's photographs. A model that repaints pixels is not enhancement, and it has no place near a news photo. It belongs in the ComfyUI lane you rent pods for.

Two seats for the bench, and neither needs a download

Scaleway put Qwen 3.8 27B on its Generative APIs. That is the local agent candidate I pointed at in August, which I have since described twice as untested. Your internet is a 5G link on a good day, and a 27B checkpoint is exactly the download you are supposed to flag. Now you can reach it on an EU provider you already have an account with, with no download at all. That also fits your standing rule of EU when quality is equal.

OpenAI's GPT-6 Luna came out at 10 cents per million input tokens and 50 cents per million output tokens. Director reran the editorial reviewer rows on the 26th. Luna is cheap enough to be one more row in that table and nothing more, and it earns the seat only if it beats the rows already there.

Heads up

Anthropic shipped Sonnet 5.5 on Monday, and the trap I described in early September has moved to where it actually hurts. Then, I said forced tool choice fails on Fable 5.1 and that Ask PärKit's natural language to SQL step was safe because it runs on Sonnet. Sonnet 5.5 now rejects forced tool choice too, and so does Opus 5.5. Your own standing instruction says to reach for the latest model. If Ask builds its structured output by forcing a tool, and that is the common way to do it, the upgrade that looks routine is the one that breaks it. Sonnet 5.5 has also removed the plain setting that switches thinking off. Checking Ask is a five-minute grep, and it should happen before anyone changes a model string.

FastMCP had a rough week. I told you to land on 4.0.5. Version 4.0.6 added a visibility check on completions. In 4.0.7, that check could let a response cache serve hidden prompts to other clients. 4.0.8 reverted it. For PärMCP, where every safeguard is a per-client gate, that is precisely the wrong kind of leak. The target is now 4.0.10, and you should never stop on 6 or 7. Also in that run: a fix to refresh token revocation in the OAuth proxy, and Google tokens removed from request URLs.

MLX core 0.32.3 fixes a Metal fence deadlock and a stream-clearing deadlock that held Python's interpreter lock. On the 23rd, the Assistant logged a Metal watchdog panic as its own class of failure. I cannot tell you these are the same bug. I can tell you the release is in the right area, and ordagrant switched to MLX-first this same week. Record which core version was running when the panic hit before you upgrade anything.

Podman shipped another security release. Running a checkpoint image could turn off sandboxing entirely, so running checkpoint images is now simply removed. Your estate does not run checkpoint images, so your exposure is close to zero. The bump costs nothing: 6.1.3, or 5.8.8 on the older line.

The rest, briefly

Pocket ID 2.16 fixes YubiKey login in some Safari versions. Floor has already bumped to it and re-frozen the restore manifest, so that one is done.

uv spent four releases on housekeeping, and its one security advisory affects Windows only. The fix worth knowing about is that a failed or interrupted upgrade now restores your project file.

R2 added per-bucket bandwidth metrics. Put next to last time's access logs, that is a second, cheaper answer to whether anyone besides you downloads the agent feeds.

PostgreSQL 19 cut several features in its fourth beta, property graph queries among them, and general availability may come in October. You are on 18 and there is nothing to do.

Xcode 27.2 beta 2 is out. You retired the beta pin three days ago, so stay retired.

ElevenLabs speech to text now takes plain-language edit instructions. Ordagrant's second invariant rules that out entirely, so this one is not for you.

If you look at one thing

Render the same two-voice passage on Eleven v4 and on Inworld, label both, and listen in the car. Credits that expire either way pay for it, and you will have an answer before the month's allowance resets.