Which IBM Granite Speech Model Fits Your Recording?

Around October 21, 2026 Beijing time (IBM’s Hugging Face blog is dated October 21, 2026; the 5.0-2b model card lists a release date of October 20), IBM refreshed its Granite Speech lineup with three pieces that look related but are not interchangeable. granite-speech-5.0-2b is new: higher-accuracy multilingual transcription and speech translation. granite-speech-5.0-470m-turboctc is an updated 470M CTC encoder for high-throughput and on-device transcription. granite-speech-4.1-2b-plus keeps the 4.1 name as a drop-in weight update tuned for who spoke and when. IBM’s own chooser is blunt: pick by job, not by “Granite Speech” as a single product.

How to choose by recording type. Need high-accuracy multilingual ASR, keyword biasing up to about 2,000 terms, language-ID prompting, or speech translation that includes English-to-Mandarin and English-to-Italian? Use granite-speech-5.0-2b. Need raw speed, edge or laptop deployment, and punctuated multilingual transcripts without a separate punctuation step? Use granite-speech-5.0-470m-turboctc. Working with meetings, support calls, or other multi-speaker audio where speaker labels and word-level timestamps matter as much as the words? Use granite-speech-4.1-2b-plus. Treating any one of these as a universal meeting-plus-translation stack will mis-set expectations.

Language caveats matter more than the version numbers. For granite-speech-5.0-2b, automatic speech recognition covers English, French, German, Spanish, Portuguese, and Japanese. Bidirectional speech translation is to and from English in those languages, plus English to Italian and English to Mandarin. That English-to-Mandarin path is a translation direction: spoken English in, Mandarin text out. It is not Chinese-audio transcription, and it does not make the 2B model a Chinese meeting tool. Turbo CTC’s transcription set is English, French, German, Spanish, Portuguese, Italian, and Japanese—it adds Italian relative to the 2B ASR list, and it is aimed at transcription throughput, not the 2B translation stack. granite-speech-4.1-2b-plus lists English, French, German, Spanish, and Portuguese; Japanese is not on that model card’s supported-language list. We have not measured word error rates, real-time factors, or VRAM on any of these three.

All three model cards state Apache 2.0. Official pages also cite Transformers (and, for the 2B path, vLLM, llama.cpp, and mlx-audio on Apple Silicon). That license and tooling note is separate from any claim about cloud compute cost or DigIntoAI benchmarks—neither is covered here.

For nearby speech and offline-audio context on DigIntoAI, see Ant Group’s Ming-3.0-tts on Artificial Analysis open TTS boards and when Gradium Phonon offline TTS beats a cloud API. Japanese-language product context that often sits next to JA ASR planning is Japan’s domestic LLMs in three routes.