Transcription, Translation, And Attribution Get Faster With Granite Speech 5.0

🔍 Read the full analysis: Transcription, Translation, And Attribution Get Faster With Granite Speech 5.0 on ThorstenMeyerAI.com

IBM has announced three updates to its Granite Speech collection: a new multilingual 2-billion-parameter model for transcription and translation, a faster multilingual version of its Turbo CTC model, and revised weights for a model focused on speaker attribution and timestamps. The releases are aimed at different speech-processing tasks, as the original analysis details, while the reported performance results are IBM’s evaluations and benchmark claims, not independent validation.

The new granite-speech-5.0-2b succeeds the 4.1 2B model at the same parameter count. IBM says it transcribes English, French, German, Spanish, Portuguese and Japanese, and supports translation to and from English for those languages. It also supports translation from English into Italian and Mandarin. The company says training used 390,000 hours of public audio, along with synthetic data for Japanese speech recognition, keyword biasing and translation.

IBM says the model is designed to handle noise, reverberation and distant microphones better than its predecessor. It accepts prompts specifying the input language and keyword lists of up to 2,000 terms to guide transcription and translation. When formatted output is requested, IBM reports capitalization F1 scores ranging from 89.7 to 99.5 across its test sets. The model is licensed under Apache 2.0 and supported in Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon.

The revised granite-speech-5.0-470m-turboctc adds Japanese, Spanish, German, French, Italian and Portuguese transcription alongside English, as well as native punctuation and capitalization. IBM reports a score of 19,870 RTFx on the Open ASR Leaderboard, compared with about 12,600 at the model’s August launch, while saying accuracy stayed at the same level. The updated granite-speech-4.1-2b-plus is aimed at longer and multi-speaker recordings: IBM reports a 24% reduction in word error rate on long-form test sets, a 7% reduction in two speaker-attribution error measures, and an average timestamp shift reduced by more than half. The company says its architecture, prompt format and API are unchanged.

Three Models, Different Speech Workloads

The releases give organizations separate options for translation, transcription speed, and speaker and timing information, rather than asking one model to serve every audio workflow. A multilingual model may suit teams working across several languages; Turbo CTC targets high-volume transcription; and the revised 4.1 model is intended for recordings where identifying speakers and aligning words to audio are important.

These features can affect how useful a transcript is after it is generated. More accurate timestamps can help align captions or locate sections of a recording, while speaker labels can make meeting and support-call records easier to review. Keyword biasing may help recognition of names, acronyms and product terminology. Those are plausible applications of the described features, but the supplied material does not quantify their impact in customer deployments.

The benchmark figures offer an initial comparison, not a guarantee for a particular organization’s recordings. Accuracy can vary with accents, language varieties, overlapping speech, specialized vocabulary and recording conditions. Buyers and developers still need to test models against audio representative of their own work and account for transcript review and operating costs.

How the Granite Models Differ

The three updates build on models intended for distinct speech tasks. IBM introduced Turbo CTC in August as an English-only, 470-million-parameter encoder built for speed. The October revision retains the same model size and adds six transcription languages, plus changes for difficult acoustics and more readable output. The new 5.0 2B model is positioned for multilingual transcription and translation, while the 4.1 2B Plus update emphasizes diarization, timestamps and long recordings.

IBM says the 5.0 2B model ranks second among open-weight ASR models on average English word error rate, reporting 5.06%. It labels this result unofficial and says it averages public English short-form test sets on the Open ASR Leaderboard as of October 15, 2026. The Turbo CTC speed comparison is also based on nine public English short-form test sets and is dated October 2026. Lower word error rate and higher RTFx measure different aspects of model performance and should not be treated as interchangeable.

IBM says all three models are released under Apache 2.0 and have native support in Transformers. It points users to model cards for evaluation, architecture and training details, and recommends pairing the models with Granite Guardian for production deployments needing additional risk detection.

“The model is the successor to granite-speech-4.1-2b at the same parameter count.”

— IBM, in its Granite Speech announcement

Benchmark Limits and Open Questions

The supplied material does not include independent validation of IBM’s reported results. Public short-form benchmark scores do not establish how the models perform across all real-world audio, and IBM’s stated results may not reflect every accent, language variety, overlapping conversation or business workflow.

Some supporting details are also missing. The capitalization F1 range for the 5.0 2B model is not broken down by test set in the supplied material. IBM describes Turbo CTC’s accuracy as unchanged while reporting higher RTFx, but the full benchmark table and the hardware and deployment conditions behind those speed results are not provided here. No customer deployment results or quantified impact on transcription costs are specified.

What Model Users Can Check

IBM directs prospective users to the model cards for fuller evaluation results, architecture and training details. Those materials should help clarify which test sets support the headline figures and how the models are intended to be used. The supplied announcement does not specify a later release date, independent test schedule or customer rollout milestone.

For teams considering adoption, the practical next step is to compare the models on representative internal audio, including the relevant languages, recording conditions and speaker patterns. That evaluation can test whether IBM’s benchmark claims carry over to a particular workflow and whether the transcript quality, timing and labeling meet its needs.

Key Questions

What did IBM announce?

IBM announced one new Granite Speech model and revisions to two existing models, targeting multilingual transcription and translation, faster transcription, and improved speaker attribution and timestamps.

Which languages does the new 5.0 2B model support?

IBM says it transcribes English, French, German, Spanish, Portuguese and Japanese. It supports translation to and from English for those languages, plus translation from English into Italian and Mandarin.

What performance gains does IBM report?

IBM reports higher Turbo CTC speed than at its August launch while maintaining accuracy, plus lower word error and speaker-attribution error measures and improved timestamps for the updated 4.1 2B Plus model. These are company-reported results; the supplied material does not provide independent validation.

Are the models open source?

IBM says all three releases use the Apache 2.0 license. The new 5.0 2B model is also listed as supported in Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon.

What should organizations verify before using them?

Organizations should test the models on audio representative of their own work, including relevant accents, terminology, languages and recording conditions. The public benchmark figures do not establish performance for every deployment.

Primary source: Hugging Face · via ThorstenMeyerAI.com