IBM has introduced the discharge of the Granite 3.3 AI mannequin lineup. Granite Speech 3.3 8B, a speech-to-text (STT) mannequin that excels in computerized speech recognition (ASR) and computerized speech translation (AST), is within the highlight.
The STT mannequin is constructed on prime of Granite 3.3 8B Instruct, a big language mannequin, with a 2B sibling model additionally out there. It options improved reasoning talents. Furthermore, its base fashions, Granite 3.3 8B Base and Granite 3.3 2B Base, are additionally out there for builders to fine-tune.
All of the fashions are launched open supply underneath an Apache 2.0 license.
Granite Speech 3.3 features a speech encoder, speech challenge, an LLM, and low-rank adaptation (LoRA) adapters.
The corporate defined that the speech mannequin is a compact and cost-efficient audio-in (and text-in), text-out STT mannequin tailor-made for enterprise use circumstances. It talked about that Granite Speech 3.3 gives better accuracy than main open and closed mannequin rivals when examined with notable public datasets.
Granite Speech 3.3 8B additionally achieved a decrease error charge for transcription duties, as indicated by the benchmark assessments.
The mannequin additionally gives automated translation from English to a various set of languages, together with French, Spanish, Italian, German, Portuguese, Japanese, and Mandarin, attaining efficiency on par with proprietary fashions like OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash on supported languages.
To assist enhance Granite-driven purposes, IBM has launched retrieval-augmented generation-focused LoRA adapters for the beforehand launched Granite 3.2 8B Instruct. These might be accessed on Hugging Face as a part of Granite Experiments.
As a part of the announcement, IBM talked about a number of areas for enchancment. At present, the audio encoder for the speech mannequin helps solely English, in order that they need to help multilingual encoding.
The corporate additional talked about different refinements, akin to knowledge recipes with higher-quality coaching knowledge and a unified construction to combine audio options in coaching levels. The corporate additionally plans to help speech emotion recognition (SER) capabilities.
The corporate talked about that it’s coaching Granite 4.0, a brand new technology of fashions that goals to have important beneficial properties in pace, context size, and capability.
The put up IBM Introduces Granite 3.3 Sequence of AI Fashions appeared first on Analytics India Journal.
