Skip to content
Cyberttak.
Learning & language· 2 min read

Granite Switch 4.1.1 swapped in a new speech model and Decision adapter without retraining the base model

IBM's patch release updates speech-to-text across 10 languages and stays fully compatible with version 4.1, with no migration.

By CyberttakPublished · 15:36

In brief

IBM patched an AI model like an app

Granite Switch 4.1.1 updates speech-to-text in 10 languages and a Decision adapter. The base model was not retrained. It is fully compatible with 4.1.

A software developer wearing headphones works at a desk in a quiet office, looking at a laptop with chat bubbles and a row of coloured dots, with a second monit
A software developer wearing headphones works at a desk in a quiet office, looking at a laptop with chat bubbles and a row of coloured dots, with a second monitCredit: AI-generated · GPT Image 2

IBM has released Granite Switch 4.1.1, a patch update to its modular language model, and the release note on the Hugging Face blog explains what changed.

It arrives just over a week after Granite Switch 4.1. That version showed how one model can carry small add-on modules called adapters. Each adapter adds a skill, and all of them run from a single endpoint, meaning one place where you send your requests.

What is new

  • An updated Granite Speech ASR model. ASR means automatic speech recognition: turning spoken words into text. It covers 10 languages and adds punctuation and capitalization natively.
  • An updated Decision adapter in granitelib-core, with better overall performance on the Jev Decision Index test set.

The punctuation detail matters in practice. Transcripts come out ready for other models to read, so nobody needs to run a separate formatting pass.

Why the update was fast

The number 4.1.1 is meant literally, as in ordinary software. It is fully backwards compatible with 4.1, and users do not need to migrate.

IBM says the base model was not touched. Only the new speech model and the Decision adapter were trained. Nothing else was recomputed.

Imagine your phone updating one app instead of reinstalling the whole system. That is the idea: pick the latest adapters you want and deploy them.

IBM also says it is working on tutorials and an Adapter SDK. The goal is for developers and third-party providers to build their own adapters and ship versions the way software ships. The company says more will come soon.

A Speech-to-Decisions demo shows the new multilingual and punctuation features.

The source gives no numbers for the Decision adapter's improvement, and it does not name the 10 languages. The Adapter SDK and tutorials are announced, not released, and no date is given.

Source: Hugging Face Blog