The Pulse
IBM Updates Granite Speech With a 19,870 RTFx Model
IBM’s Granite Speech update includes a Turbo CTC model reported at 19,870 RTFx and a multilingual 2B model with an official 5.06% average English word error rate. The release also updates a model for speaker attribution, timestamps, long re
AI.info Team · Written with AI · fact-checked ·
IBM says its updated Granite Speech Turbo CTC model reaches 19,870 RTFx on the Open ASR Leaderboard, processing audio far faster than real time. The October 21 announcement also introduces a 2-billion-parameter model for multilingual transcription and translation, and refreshes a model designed to identify speakers and mark word timings. IBM is positioning the three models for different speech tasks, rather than offering one model for every transcript workflow.
19,870 RTFx, with a leaderboard caveat
Turbo CTC is a 470-million-parameter model built for high-throughput and on-device transcription. Its update adds Japanese, Spanish, German, French, Italian and Portuguese to English, and produces punctuation and capitalization without a separate processing step. IBM says a change to how it compiles and reuses a Conformer block raised its reported speed from about 12,600 RTFx at launch to 19,870 RTFx.
The speed figure needs qualification: the updated model card labels Turbo CTC’s leaderboard result unofficial because the leaderboard entry still corresponds to the earlier English-only release. IBM also reports that the model’s average English word error rate improved from 5.78% to 5.51% after the update. Those figures come from IBM’s tests, not an independent comparison of real customer recordings.
The 2B model adds translation and keyword lists
The new Granite Speech 5.0 2B succeeds Granite Speech 4.1 2B at the same parameter count. Its model card lists October 20, 2026, as its release date. It transcribes English, French, German, Spanish, Portuguese and Japanese, and translates between English and those languages; it also translates from English into Italian and Mandarin. IBM says the model was trained on 390,000 hours of public audio along with synthetic data for Japanese recognition, keyword biasing and translation.
Users can specify the input language in a prompt and supply keyword lists of up to 2,000 terms, such as product names, people and acronyms. IBM reports that the model ranks second among open-weight speech-recognition models on average English word error rate, with an official score of 5.06% across public short-form test sets as of October 21, 2026. Lower error rates indicate better transcription.
The announcement’s authors George Saon and Brian Kingsbury write:
Translation holds up for longer with more natural speech, not just short read sentences.
The model is released under Apache 2.0 and supports Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon. The model card advises transcribing in segments of up to about two minutes and splitting longer recordings before decoding.
Speech 4.1 Plus targets meetings and calls
The third update refreshes Granite Speech 4.1 2B Plus, which is aimed at meetings, customer-support calls and interviews with multiple speakers. It keeps the same architecture, prompt format and API, making the new weights a drop-in replacement, IBM says. The update focuses on speaker attribution, timestamps, long recordings and keyword recognition.
In IBM’s reported results, word error rate on long-form test sets falls from 13.53 to 10.22, while errors on biased keywords with 20-term lists drop from 11.79 to 6.48. Average timestamp shift for all speech declines from 210.86 to 98.54, and IBM reports 7% reductions in both word diarization error rate and speaker-attributed word error rate overall. These measures address practical transcript problems—misheard names, incorrect speaker labels and timing that drifts from the audio.
Three models, three deployment choices
All three models are available under Apache 2.0, according to IBM’s announcement, which also lists native Transformers support. The 2B model is the option for multilingual transcription, translation and large keyword lists; Turbo CTC is aimed at speed and edge use; and 4.1 2B Plus adds speaker labels and timestamps for longer, multi-speaker recordings. The model cards provide usage instructions and evaluation details, but the reported benchmark results do not settle how the models will perform on each organization’s own accents, microphones, background noise or terminology.