Granite Speech 3.3 8B

Architecture: Two-pass speech-language model:

1. Conformer acoustic encoder with block attention, self-conditioning and CTC training.
2. Windowed Q-Former speech-text adapter with 2 layers and 3 trainable queries per block (10× temporal downsampling).
3. Granite 3.3 8B Instruct LLM (128k context) as the language backbone.
4. LoRA adapters (rank 64) on the LLM's query/value projections, activated only in speech mode.

Modality: Audio in / text out. Supports a pure-text fallback that calls the underlying Granite LLM directly, preserving full text capabilities and safety alignment.

Training: Modality-aligned from granite-3.3-8b-instruct on public ASR/AST corpora plus synthetic translation data. This particular checkpoint was trained for 13 days on 32 H100 GPUs on IBM's Blue Vela supercomputer. Revision 3.3.2 introduced multilingual support for English, French, German, Spanish and Portuguese.

Use cases: Enterprise-grade English ASR (outperforms several competitors trained on far more proprietary data), English↔X speech translation for major European languages; suited for downstream Q&A and summarization via the Granite LLM.