IBM releases free speech-to-text models that transcribe and translate

IBM has released three updates to its Granite Speech models, which turn recorded speech into text. They are free to use under an Apache 2.0 license. The headline model, granite-speech-5.0-2b, handles six languages and can translate to and from English. It ranks second among top open models for English accuracy.

What changed

The new model, granite-speech-5.0-2b, replaces granite-speech-4.1-2b at the same size. It transcribes English, French, German, Spanish, Portuguese and Japanese. It translates to and from English in those languages, and also from English into Italian and Mandarin. IBM says it trained on 390,000 hours of public audio, plus synthetic data for Japanese, keyword focus and translation.

On the Open ASR Leaderboard's public English short-form tests, it scored a 5.06% average word error rate. Lower is better. IBM lists that as second among top open models as of October 21, 2026.

Two older models got updates

• Turbo CTC, a small 470M-parameter model built for speed, started in August as English-only. This revision is now multilingual and more robust.
• Turbo CTC is also about 1.6x faster, at 19,870 on a speed score, up from roughly 12,600. IBM says accuracy stayed at the same level.
• granite-speech-4.1-2b-plus is tuned for meetings, support calls and long interviews with several speakers, where who spoke and when matters.

The 4.1 Plus update keeps its name because its design and interface are unchanged. IBM calls it a drop-in replacement with improved weights, meaning the learned settings inside the model. IBM says it improves four areas, but the details are on the model card.

Who it affects

Businesses sitting on call or meeting recordings now have free options for building transcripts. The Apache 2.0 license is permissive, so the models can go into commercial products. They work with common tools including Transformers, vLLM and llama.cpp. A version for Apple Silicon is supported through mlx-audio.

What is not clear yet

The accuracy figures come from IBM's own report of public tests, not from recordings like yours. Real calls with noise, accents or jargon may score differently. IBM also recommends pairing the models with its Granite Guardian tool for production use that needs extra risk detection. Full results, training data and usage examples are on each model card.