IBM Unveils Granite Switch 4.1, Bundling Conversation and Speech Recognition
One model handles conversation, speech transcription, and harmful content screening. IBM says these capabilities use roughly the same amount of GPU memory as the base model.
IBM has released Granite Switch 4.1, a model that bundles conversation, decision-making, and speech recognition into a single package. Two versions are available: Granite-Switch-4.1-8B and Granite-Switch-4.1-30B. Each takes a base model and attaches purpose-built capabilities on top, so that several functions can be called from one place.
Three capabilities are included: Decision, which handles general-purpose decision tasks; Guardian, which specializes in safety-related decisions; and Granite-Speech-5.0-Turbo-CTC, which turns speech into text. IBM says this speech recognition model is the fastest on the OpenASR leaderboard, which compares speech recognition performance.
Do more capabilities mean more GPUs?
The change IBM is highlighting is that adding capabilities keeps GPU memory usage roughly the same as running the base model alone. The company says this setup lets a smaller number of GPUs handle multiple functions. For services that run conversation, speech processing, and content screening together, this has direct implications for hardware requirements.
Image: huggingface.co
The mechanism is the “adapter,” a small additional training module that bolts a capability onto the base model. Information computed during a conversation is reused in the decision tasks that follow. IBM says software optimizations that speed up processing, combined with integrated speech recognition, allow fast switching between conversation, decision, and speech requests.
This approach of composing capabilities also appeared in the earlier Granite Libraries. In that case, adapters tuned for specific decisions or tasks were collected in one place and users picked and combined them. The two models released this time come as pre-assembled bundles of useful adapters. Users can also go the other route, selecting only the adapters they need from Granite Libraries to build their own model.
Beyond transcription, how far does it go?
The use case IBM points to is transcribing speech, summarizing the content, and then checking whether anything harmful is in it. IBM says a single 8B model handles these functions along with 15 user-defined screening criteria in real time. The work does not stop at converting speech to text; it continues into organizing and inspecting that content.
Image: huggingface.co
In performance comparisons, IBM leaned on the benefits of task-specific capabilities. According to the company, Guardian scored higher on average than Jev 1.13, Microsoft-Decision-1, GPT-Luna-Decisions, and Clef-Flash on safety-related decision tasks. The 30B model with four specialized adapters attached completed a five-step task that answers questions by referring to retrieved material far faster than Jev-27B, IBM said.
IBM plans to release additional developer tools and guidance so users can train their own adapters. In the end, the question is how easily you can attach and combine capabilities to match the work you actually need done.