Granite Speech 3.3

Granite Speech 3.3 (2025, arXiv 2505.08699) is IBM's first open speech-to-text and speech-translation model, built by modality-aligning Granite 3.3 Instruct (2B and 8B) to speech: a conformer acoustic encoder with block attention, a windowed query-transformer adapter that maps acoustic embeddings into the LLM's text space, and LoRA adapters for the final tuning. The practical pitch is enterprise transcription and translation: IBM reports it beats competitors trained on far more proprietary data on English ASR, keeps pace on English→X translation for French, German, Spanish, Portuguese, Japanese, Italian, and Mandarin, and — unlike Whisper's 30-second windows — accepts arbitrarily long audio (a 20-minute file in one go). Both sizes are on Hugging Face under Apache 2.0, with GGUF community quantizations for CPU inference. If you want an open, license-clean speech translation backbone you can fine-tune, this is one of the strongest options published.

• English ASR outperforming models trained on far more proprietary data (author-reported)\n- Competitive English→X translation for major European languages, Japanese, Chinese (author-reported)\n- No independent benchmark replication found