Granite Speech 3.3 2B

Architecture: A two-pass speech-language model composed of:

1. Conformer acoustic encoder using block attention with self-conditioning, trained with Connectionist Temporal Classification (CTC).
2. Speech projector / temporal downsampler — a 2-layer windowed Q-Former operating on blocks of 15 1024-dim acoustic embeddings, producing a 10 Hz embedding rate (10× total temporal downsampling).
3. Granite 3.3 2B Instruct LLM backbone with 128k context length.
4. LoRA adapters (rank 64) on the LLM's query and value projection matrices, activated only in speech mode.

Modality: Speech-to-text. Operates in two modes: speech mode (encoder + projector + LoRA active for ASR/AST) and text mode (pure Granite 3.3 LLM, preserving safety and text capabilities).

Training: Trained on IBM's Blue Vela cluster (NVIDIA H100) using publicly available ASR/AST corpora plus synthetic data targeted at the speech-translation task. Revision 3.3.2 added multilingual inputs (English, French, German, Spanish, Portuguese) and a deeper acoustic encoder for improved English ASR.

Use cases: Enterprise transcription, English and European-language ASR, English↔X speech translation, and as a building block for downstream Granite text workflows.