Granite Speech 5.0

IBM's Apache 2.0 speech-to-text pair: a 2B model that transcribes and translates, and a 470M CTC model built for speed and edge devices.

In short. Granite Speech 5.0 is IBM's pair of speech-to-text models. The 2B model transcribes six languages and translates speech to and from English. The much smaller 470M model is built for speed and can run on a laptop or phone. Both can be used commercially.

Summary

IBM released two open speech models under the Granite Speech 5.0 name, and its blog post announcing them is dated 21 October 2026. granite-speech-5.0-2b, released on 20 October, is a speech-language model of about 2.3B parameters. It combines a 16-block conformer speech encoder, a speech projector and a Granite 4.0 1B language model. It transcribes English, French, German, Spanish, Portuguese and Japanese, translates to and from English in those languages (plus English to Italian and Mandarin), accepts a language-ID prompt, and can be biased toward up to 2,000 keywords. IBM trained it on 390,000 hours of public audio plus synthetic data. granite-speech-5.0-470m-turboctc is a 470M-parameter conformer with a CTC output layer, so it decodes without generating one token at a time. Its English-only original appeared on 25 August, and on 21 October IBM updated it to seven languages (the six above plus Italian) with native punctuation and capitalization.

IBM says the 2B model ranks second among open-weight models on the Open ASR Leaderboard's English short-form average, with a 5.06% word error rate in official results as of 21 October 2026. It says TurboCTC averages 5.51% on English and reaches 19,870 RTFx on the same leaderboard, about 1.6 times faster than at launch. These are IBM's own figures read from a public leaderboard that changes often. For the 2B model the card recommends splitting recordings longer than about two minutes before decoding.

Both models are Apache 2.0. The BF16 weights are about 5.0 GB for the 2B model and 0.95 GB for TurboCTC. Transformers supports the 2B model from 4.52.1 and TurboCTC from 5.16.0, and the blog lists vLLM, llama.cpp and mlx-audio support for the 2B model, plus browser demos using WebGPU. Two related models are separate. granite-speech-5.0-470m-turboctc-nc is English-only and licensed CC-BY-NC-SA 4.0 for research and non-commercial use, and granite-speech-4.1-2b-plus is IBM's pick for meetings with several speakers. This entry covers only the two Apache 2.0 Granite Speech 5.0 models.

Good for

  • Accurate transcription in English, French, German, Spanish, Portuguese and Japanese, plus translation to and from English with the 2B model
  • On-device or high-volume transcription with the 470M TurboCTC model, which IBM says reaches 19,870 RTFx
  • Transcripts that must get product names and acronyms right, since the 2B model accepts keyword lists of up to 2,000 terms
  • Commercial products, since both models use Apache 2.0

Watch out for

  • Chinese, Korean, Arabic and Hindi transcription is not covered, and the 2B model only translates English to Mandarin
  • IBM points to Granite Speech 4.1 2B Plus, not these models, for meetings with several speakers
  • The 2B model works best on segments of up to about two minutes, so long recordings must be split first
  • granite-speech-5.0-470m-turboctc-nc is non-commercial, so check the repository name before you use it
  • Word error rates are IBM's own and come from a leaderboard that changes

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Granite Speech 5.0. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Granite Speech 5.0 470M TurboCTC BF16 470M 0.95 GB 1.09 GB Laptop or phone
Granite Speech 5.0 2B BF16 2.31B 5.04 GB 5.8 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Granite Speech 5.0?
Granite Speech 5.0 is IBM's pair of speech-to-text models. The 2B model transcribes six languages and translates speech to and from English. The much smaller 470M model is built for speed and can run on a laptop or phone. Both can be used commercially.
Can I use Granite Speech 5.0 commercially?
Yes. Granite Speech 5.0 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Granite Speech 5.0 open source?
Granite Speech 5.0 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Granite Speech 5.0 on my own computer?
Yes. The smallest variant, Granite Speech 5.0 470M TurboCTC, needs about 1.09 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Granite Speech 5.0 released?
Granite Speech 5.0 was released by IBM on October 20, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Voxtral Realtime Mistral AI 4.43B 128K Apache 2.0 (Permissive) 10 GB Laptop or phone
Qwen3-ASR Alibaba Qwen 918M – 2.35B – Apache 2.0 (Permissive) 2.12 GB Laptop or phone
Nemotron 3.5 Transcribe NVIDIA 34B 4.3B active 64K OpenMDW-1.1 (Permissive) 25 GB Gaming GPU
Whisper large-v3-turbo OpenAI 810M – 1.54B – MIT (Permissive) 6 GB Laptop or phone