Evidence at a glance
Three Updates for Three Different Pressures
In October 2026, IBM Granite announced three updates to its Granite Speech collection: the new granite-speech-5.0-2b, a revised granite-speech-5.0-470m-turboctc, and new weights for granite-speech-4.1-2b-plus. They target three different enterprise speech problems: multilingual transcription and translation, transcription at very high speed, and identifying who said what and when in recordings with multiple speakers.
The interesting part is that IBM has not presented these as one model becoming better at everything. A speech system does more than produce a string of words: translation may matter more than verbatim transcription, latency can determine whether a live workflow is usable, and meeting records need speaker labels and timestamps. The lineup acknowledges that these goals do not always fit one model or one score.
Different Architectures, Different Trade-offs
granite-speech-5.0-2b retains a 2-billion-parameter scale and uses a speech encoder, a projector, and a language model for recognition and translation. Prompts can specify the input language and task, and can include up to 2,000 keywords to guide recognition of names, product terms, or acronyms. The model transcribes English, French, German, Spanish, Portuguese, and Japanese; it translates between those languages and English, and also supports translation from English into Italian and Mandarin. IBM says it was trained on 390,000 hours of public audio, alongside synthetic data for Japanese recognition, keyword biasing, and translation.
Turbo CTC takes a different route. It is a 470-million-parameter Conformer encoder using CTC and non-autoregressive greedy decoding, designed to transcribe quickly rather than to bring every generative language-model capability into the task. The revised version adds English, Japanese, Spanish, German, French, Italian, and Portuguese, along with native punctuation and capitalization. IBM also reports compiling one Conformer block and reusing its computation graph across the stack to increase throughput. These are not simply large and small versions of the same model: the 2B model offers more prompt-based control over tasks and terminology, while Turbo is built around a faster processing path.
Meeting Records Need More Than Correct Words
The third update, granite-speech-4.1-2b-plus, keeps its architecture but strengthens speaker attribution, word-level timestamps, and keyword biasing for long recordings with multiple speakers. IBM reports overall reductions of 7% in both word diarization error rate (WDER) and speaker-attributed word error rate (cpWER), a reduction of more than half in average timestamp offset, and a 24% decrease in long-audio WER. These metrics measure different things: correct words do not guarantee correct speaker labels, and an accurate sentence does not tell a reviewer exactly when each word was spoken.
That distinction affects whether a transcript can flow directly into downstream work. Meeting search needs to connect content with speakers, review needs to locate the relevant moment, and support-call analysis may depend on turn boundaries. Keyword biasing is another business-facing control: IBM says keyword errors fell by 45% when the model used a 20-word list. That does not mean prompts can solve every acoustic problem, but it gives organizations a way to make specialized terms less dependent on a general-purpose model having encountered them during training.
What the Benchmark Numbers Do—and Do Not—Show
IBM reports an average word error rate (WER) of 5.06% for the 5.0-2b on the Open ASR Leaderboard's public English short-form test sets, placing it second among open-weight ASR models as of October 15, 2026. This is useful but bounded evidence: it offers a comparison for short-form English recognition, not an overall measure of multilingual translation, far-field noise handling, or long meetings. IBM labels the result unofficial.
Turbo CTC's numbers more directly reflect its product goal. IBM reports 19,870 RTFx on the same leaderboard, up from roughly 12,600 at launch, or about 1.6 times faster, while maintaining the same accuracy level. The model card also gives an average English WER improvement from 5.78% to 5.51%. RTFx measures processing speed and WER measures recognition errors; neither substitutes for the other. Higher throughput matters for teams processing large volumes of recordings, but it does not establish that the model is more accurate across every language and acoustic condition.
Choose by Workflow, Not by a Single Score
A practical evaluation can start with the required output: assess 5.0-2b for multilingual recognition, translation, or terminology biasing; Turbo CTC when transcription throughput is the priority; and 4.1-2b-plus when meetings or calls depend on speaker labels, word-level timestamps, and long-audio performance. This is not a unified router supplied by IBM, but an engineering selection approach inferred from the models' stated capabilities and goals. If a system needs all of these, test each stage separately rather than assuming a leaderboard rank covers the entire pipeline.
Integration costs differ too. The 5.0-2b ships under Apache 2.0 and supports Transformers, vLLM, llama.cpp, and mlx-audio. Existing users of 4.1-2b-plus can replace the weights without changing the architecture, prompt format, or API. The final decision should still be based on internal data: test the real language mix, far-field recordings, domain vocabulary, and long multi-speaker audio, measuring WER, speaker errors, timestamp offset, and throughput. The published material does not provide complete evaluations by language and noise condition for those scenarios. The safer conclusion is that the models now have clearer roles, not that broad production superiority has already been demonstrated.