IBM updates Granite Speech: new 5.0-2B multilingual model, Turbo CTC about 1.6x faster
According to an official IBM blog post published on October 21, 2026, the Granite Speech series has three updates: the new model granite-speech-5.0-2b supports six-language transcription and bidirectional translation with English; the Turbo CTC 470M version adds multilingual support and improves RTFx by about 1.6x; and 4.1-2b-plus improves speaker attribution and timestamps. All three are released under the Apache 2.0 license.
What changed
According to an official IBM blog post published on October 21, 2026, the Granite Speech series has three updates. The new granite-speech-5.0-2b is a same-parameter-count successor to granite-speech-4.1-2b and can transcribe English, French, German, Spanish, Portuguese, and Japanese, and translate between these languages and English; it also supports English-to-Italian and English-to-Mandarin translation. Training used 390,000 hours of public audio plus synthetic data built for Japanese ASR, keyword biasing, and translation. New capabilities include language ID prompting, keyword bias lists of up to 2,000 words, and prompt-driven punctuation and capitalization output (the blog says capitalization F1 on its test sets is 89.7 to 99.5). The released granite-speech-5.0-470m-turboctc revision keeps the 470M encoder size, adds transcription for English, Japanese, Spanish, German, French, Italian, and Portuguese, and natively outputs punctuation and capitalization; IBM says that by compiling a single Conformer block and reusing the computation graph across the stack, the model reaches 19,870 RTFx on the Open ASR Leaderboard versus about 12,600 RTFx at release, roughly 1.6x faster at the same accuracy level. The revised granite-speech-4.1-2b-plus keeps the 4.1 name; IBM says the architecture, prompt format, and API are unchanged, making it a direct drop-in weight improvement. The blog lists word-level speaker diarization error rate, WDER, dropping from 4.40 to 4.10; WER with speaker attribution dropping from 30.50 to 28.48; average timestamp offset across all speech dropping from 210.86 to 98.54; long-form WER dropping from 13.53 to 10.22; and biased word WER with a 20-word keyword list dropping from 11.79 to 6.48. The blog says all three models are released under the Apache 2.0 license, are natively supported in Transformers, and that 5.0-2b also supports vLLM, llama.cpp, and mlx-audio for Apple Silicon. The above numbers and rankings are all self-reported by IBM.
What it means for developers
The direct impact for developers centers on three alternative replacement paths rather than a single model upgrade. First, if an existing pipeline uses a speech language model for multilingual transcription and is sensitive to proper-noun errors, the prompt-level controls introduced by 5.0-2b are a new evaluable variable: language ID prompting, keyword biasing up to 2,000 words, and on-demand punctuation and capitalization are all inference-time inputs rather than weight-layer changes, so controlled tests can be run without changing the deployment architecture. Second, the Turbo CTC revision adds multilingual support, native punctuation and capitalization, and training for reverberant audio, while IBM reports RTFx rising from about 12,600 to 19,870, which for edge or high-throughput scenarios implies a significant increase in the audio duration processable per unit of hardware; but RTFx is a throughput metric, and IBM says the improvement is achieved at the same accuracy level, so it cannot be used to infer error rates under specific noise and accent distributions. Third, the 4.1-2b-plus revision is described as a direct drop-in with unchanged architecture, prompt format, and API, and shows the largest improvements in timestamps, long-form speech, and keyword biasing, which could reduce review and subtitle alignment costs in meeting and customer-service recording pipelines, but replacements still need to be validated on your own audio. Note that these models are open weights; the Apache 2.0 license stated in the blog and the language coverage for transcription and translation should be treated as a paraphrase of the model card and the publisher, and downloadable weights do not automatically equal full open-source license rights; commercial use and redistribution terms should still be checked against the accompanying license text. In addition, these three updates are not about generic speech infrastructure; they are specific model version changes, but their API and service availability are determined by the publisher and are unrelated to any third-party model catalog.
Integration and evaluation
The following is conditional analysis and advice for developers, not verified fact. Decision one: whether to switch multilingual transcription to granite-speech-5.0-2b. The basis is the new capabilities reported by the source: language ID prompting and keyword biasing up to 2,000 words. When making the choice, distinguish whether the input language is clearly known in the prompt: if the language can be determined per item, language ID prompting can be included in inference requests and compared against a no-prompt baseline; if the audio mixes languages or the language is unknown, the benefit of that prompt is uncertain, and its automatic language detection performance should be the focus. The prerequisite is that the inference stack supports prompted input (the blog says Transformers has native support and lists vLLM, llama.cpp, and mlx-audio). A measurable evaluation criterion: on a keyword list of about 20 to 200 words made of your own product names, person names, and acronyms, compare the relative change in biased word error rate with biasing enabled versus disabled, and also record whether overall WER degrades. Decision two: whether to replace the original English version with the Turbo CTC 470M revision for edge or high-throughput transcription. The basis is the two changes reported by the source: added multilingual support and native punctuation and capitalization, plus IBM's reported RTFx improvement. The condition for choosing is whether the business needs non-English transcription: if only English is needed and the existing pipeline already has an external punctuation restoration step, the main benefits of replacement are removing that step and higher throughput; if multilingual is needed, first confirm that the target languages are among the seven listed in the blog. Evaluation criterion: under fixed concurrency and hardware, measure processing time per unit duration of audio and the actual word error rate, and separately count samples with reverberation or far-field recordings, because the blog says this version is explicitly trained for reverberant audio. Decision three: whether to use 4.1-2b-plus as a direct replacement for meeting and multi-speaker recordings. The basis is the reported weight improvements and the unchanged items: architecture, prompt format, and API are unchanged, and WDER, cpWER, all-speech timestamp offset, and long-form WER all decreased. The condition for choosing is that the existing pipeline is already connected to the 4.1 prompt and interface, in which case the replacement cost is mainly regression validation rather than code changes. Evaluation criterion: on your own long meeting recordings, count both speaker attribution error rate and timestamp offset, focusing on speaker-change points and drift in the later part of long audio, because the improvements given by the source are not consistent across subsets. All of the above evaluations should be done on your own data; the source's self-reported metrics should only serve as a prior.
Limits and open questions
The above metrics, rankings, and speed data all come from IBM's self-reported blog and model card; this brief has not independently reproduced them or verified them on its own data. Several specific numbers lack comparable third-party reproduction: for example, the comparison of 19,870 RTFx versus about 12,600 RTFx comes from the same publisher's measurement methodology and is not directly comparable across environments; the capitalization F1 range of 89.7 to 99.5 does not explain how each test set was split. The source does not disclose per-language error rates for the new languages in the Turbo CTC revision or the 4.1-2b-plus revision, nor does it give broken-down metrics for each language's bidirectional translation with English, so there is significant unknown in multilingual performance. The source also does not explain the accuracy of keyword biasing at the 2,000-word limit or the latency cost of very long keyword lists. It should be clearly stated that none of these three versions currently have a corresponding entry in KimiSeek's live catalog; the model availability and access methods described in this brief (Transformers, vLLM, llama.cpp, mlx-audio, etc.) refer to IBM's release channels, not to this platform having them listed. Whether they are later listed and the corresponding call pricing should be based on KimiSeek's pricing page. On licensing, the blog says they are released under Apache 2.0, but downloadable weights do not equal complete license rights; actual commercial use and redistribution terms still need to be checked against the accompanying license and model card. Finally, the source recommends pairing with Granite Guardian for risk detection in production deployments; that recommendation itself comes with no evaluation data, and whether to adopt it requires your own testing.