IBM Granite Speech
Open-weight Apache 2.0 speech recognition and translation models for enterprise transcription.
About IBM Granite Speech
IBM Granite Speech: Open-Weight Speech Recognition and Translation
Granite Speech is a family of open-weight speech-language models for automatic speech recognition (ASR) and speech translation in English, French, German, Spanish, Portuguese and Japanese. All models ship under the Apache 2.0 license on Hugging Face, so teams can run, fine-tune and deploy them commercially on their own infrastructure.
Key Features
Granite Speech 5.0 2B: the flagship 2B-parameter model for high-accuracy transcription and translation, with a 5.06% average English word error rate on the Open ASR Leaderboard's public test sets (October 2026).
Turbo CTC 470M: a compact, high-throughput model for edge and on-device transcription in seven languages, with native punctuation and capitalization.
Speaker attribution and timestamps: Granite Speech 4.1 2B Plus labels who spoke and when, built for meetings, support calls and interviews.
Keyword biasing: pass up to 2,000 product names, people or acronyms in the prompt to improve recognition of domain terms.
Speech translation: translates to and from English across supported languages, plus English to Italian and Mandarin.
Language ID prompting: name the input language in the prompt to steer multilingual transcription.
Formatted output on request: punctuation and capitalization can be switched on through the prompt.
Robust audio handling: trained to handle noise, reverberation and distant microphones.
Broad runtime support: native in Hugging Face Transformers, with vLLM (OpenAI-compatible server), llama.cpp GGUF builds and mlx-audio for Apple Silicon.
Use Cases
Self-hosted transcription: enterprise call, meeting and media transcription where audio stays on internal infrastructure.
Multilingual customer support: transcribe and translate calls across European languages and Japanese.
Meeting and call analytics: speaker-attributed transcripts with timestamps for search, review and captions.
On-device and browser speech: run the Turbo CTC model at the edge or in the browser through WebGPU.
Getting Started
Website: https://huggingface.co/collections/ibm-granite/granite-speech
Model card: https://huggingface.co/ibm-granite/granite-speech-5.0-2b
Browser demo: https://huggingface.co/spaces/ibm-granite/granite-speech-webgpu
IBM Granite Speech gives developers and enterprises a family of open, commercially usable speech models that cover high-accuracy transcription, fast edge inference and speaker-attributed meeting transcripts, all deployable on their own hardware.
