IBM announced updates to its Granite Speech collection on its official Hugging Face blog on October 21, 2026. The new granite-speech-5.0-2b model card lists October 20 as its release date; that is distinct from the announcement date. This is speech recognition and speech-to-text translation, not text-to-speech generation. The announcement also covers multilingual TurboCTC and revised speaker attribution and timestamps in the existing 4.1 2B Plus model.
Choose accuracy or transcription speed deliberately
The 2B model supports language-ID prompting and keyword lists of up to 2,000 terms. IBM reports better handling of difficult acoustics and longer conversational translation. Its stated ASR languages are English, French, German, Spanish, Portuguese and Japanese. The faster 470M TurboCTC also includes Italian. Neither stated list includes Korean. IBM's rankings and throughput figures refer to particular evaluation sets and execution conditions, not independently verified superiority across languages and devices.
Open weights do not make operation free
Apache 2.0 weights permit self-hosted deployment. The 2B card documents Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon. Local operation can reduce the need to send recordings to an external service, but does not automatically ensure privacy or eliminate hardware, storage, electricity and maintenance costs. Consent and access controls still matter. IBM recommends splitting 2B audio into segments of about two minutes or less. Before adoption, test actual microphones, noise and names for omissions and substitutions; compare the Plus model separately when speaker identity and timing are essential.