IBM has added a new flagship to its open speech lineup. On October 21, 2026 the Granite team published granite-speech-5.0-2b on Hugging Face, together with revised versions of two models already in use: granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. All three carry an Apache 2.0 license and work natively in Hugging Face Transformers, and IBM positions them as a set covering accurate transcription and translation, very fast transcription at the edge, and multi-speaker recordings.
The 2B model succeeds granite-speech-4.1-2b at the same size class; its model card lists about 2.3 billion parameters and a granite-4.0-1b-base language backbone. Its transcription covers six languages (Japanese, Portuguese, Spanish, German, French and English); it also translates between English and those languages, and also translates from English into Italian and Mandarin. According to IBM, the training mix combined some 390,000 hours of publicly available recordings with synthetic material generated for Japanese recognition, keyword biasing and translation.
Compared with its predecessor, IBM lists better English accuracy and more tolerance of noise, echo and distant microphones; prompts that name the input language to steer multilingual transcription; stronger translation of long, conversational speech; keyword lists of up to 2,000 terms to favor product names, people and acronyms; and punctuation and capitalization when requested. IBM reports an average English word error rate of 5.06% on the Open ASR Leaderboard's public short-form sets, which it says places the model second among top open-weight recognizers as of release day. Beyond Transformers, IBM says the model runs in vLLM and llama.cpp, and on Apple Silicon through mlx-audio.
The Turbo CTC model, a 470-million-parameter encoder that IBM introduced in August as an English-only speed play, now handles seven languages and outputs punctuation and capitalization natively. IBM says training aimed at reverberant audio makes it sturdier in echoey rooms, and that compiling one Conformer block once and reusing that graph across the stack raised throughput on the leaderboard to 19,870 RTFx from about 12,600 at launch, at similar accuracy. The same Turbo CTC family already backs the speech component of IBM's Granite Switch 4.1 endpoint, released earlier in October.
The third model, granite-speech-4.1-2b-plus, keeps its name because its architecture, prompt format and API are unchanged; only the weights are new, making it a drop-in swap. IBM targets it at meetings, support calls and interviews. In IBM's tables, word diarization error falls from 4.40 to 4.10, speaker-attributed WER from 30.50 to 28.48, long-form WER from 13.53 to 10.22, and average timestamp shift from about 211 to 99.
For developers, the practical draw is permissively licensed speech models that run in common open inference stacks, with a size spread from on-device to server use. Enterprises working on call analytics or meeting tools get speaker attribution and keyword biasing without a proprietary API, and IBM recommends pairing the models with Granite Guardian for added risk detection in production.
The benchmark figures and improvement percentages are IBM's own measurements, and the leaderboard ranking is a snapshot from release day that can shift as other models are submitted. The multilingual and translation results are summarized rather than fully broken down in the announcement, so teams should check the model cards and test on their own audio.