🔍 Read the full analysis: 3 Ways Granite Speech 5.0 Speeds Up Transcription, Translation, And Attribution on ThorstenMeyerAI.com
TL;DR
IBM announced one new model and updates to two existing models in its Granite Speech collection. The releases target multilingual transcription and translation, faster transcription, and improved speaker attribution and timestamps; performance results cited in the announcement are IBM’s evaluations, not independent validation.
IBM announced one new model and revisions to two existing models in its Granite Speech collection, targeting multilingual transcription and translation, faster speech recognition, and improved speaker labeling and timestamps. The releases give organizations separate options for different audio workloads, as described in the original analysis, while the performance figures cited so far come from IBM’s evaluations and public benchmark comparisons.
The new granite-speech-5.0-2b is a 2-billion-parameter model and 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, as well as translation from English into Italian and Mandarin. The company says it trained the model on 390,000 hours of public audio, with synthetic data for Japanese speech recognition, keyword biasing and translation.
IBM says the model is better than its predecessor at handling noise, reverberation and distant microphones. It accepts prompts identifying 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 released under Apache 2.0 and is supported in Transformers, vLLM, llama.cpp and mlx-audio for Apple Silicon.
The updated granite-speech-5.0-470m-turboctc adds transcription in Japanese, Spanish, German, French, Italian and Portuguese alongside English, plus native punctuation and capitalization. IBM reports a speed result of 19,870 RTFx on the Open ASR Leaderboard, compared with about 12,600 at its August launch, while saying accuracy remained at the same level. Separately, revised weights for granite-speech-4.1-2b-plus are reported to reduce word error rate by 24% on long-form test sets, cut two speaker-attribution error measures by 7%, and more than halve average timestamp shift across speech. IBM says that update leaves the architecture, prompt format and API unchanged.
Three Models, Three Audio Workloads
The releases reflect distinct demands within speech processing. Translation across languages, high-volume transcription and recordings with multiple speakers can require different capabilities, so a single speed or accuracy figure does not describe every use case. IBM positions the 5.0 2B model for multilingual transcription and translation, Turbo CTC for speed, and the revised 4.1 2B Plus for speaker attribution, timestamps and longer recordings.
These capabilities may affect what teams can do with transcripts after recognition. More accurate timestamps can help align captions or locate passages in recordings; speaker labels can make meeting and support-call records easier to review; and keyword biasing may help with names, acronyms and product terms. These are plausible uses of the features described by IBM, not measured outcomes from specific customer deployments in the supplied material. Organizations will still need to test the models on audio representative of their own work.
The results also have practical limits. Word error rate, RTFx, speaker-attribution error and timestamp shift measure different aspects of performance; they should not be treated as interchangeable. Public benchmark results can help users compare systems on stated tests, but do not establish how a model will perform across every accent, language variety, microphone or business workflow. Deployment decisions also involve review procedures and the cost of processing audio.
As an affiliate, we earn on qualifying purchases.
How the Granite Releases Differ
IBM introduced the 470-million-parameter Turbo CTC model in August as an English-only encoder designed for speed. The October revision keeps its model size while adding six transcription languages and changes for punctuation and capitalization. The newly announced 5.0 2B model is the collection’s multilingual transcription and translation option, while the 4.1 2B Plus revision focuses on diarization—identifying who spoke when—and timestamp accuracy.
IBM says the 5.0 2B model placed second among open-weight ASR models on average English word error rate, with a reported score of 5.06%. The company describes this as 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 uses nine public English short-form test sets and is dated October 2026. IBM says all three models are available under Apache 2.0 and have native support in Transformers. It points users to model cards for evaluation, architecture and training details.
“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 reported results have not been independently validated in the supplied material. IBM’s leaderboard figures are tied to specified public test sets, and do not establish performance across all accents, overlapping speech, specialized terminology or recording environments. The announcement material also does not provide the full benchmark tables needed to inspect every comparison.
Some details behind the numbers remain unspecified. IBM reports Turbo CTC’s RTFx result but the supplied information does not identify the hardware and deployment conditions used to produce it. The capitalization F1 range for the 5.0 2B model is not broken down by test set here. It is also unclear how the claimed gains translate to particular customer workflows or what review burden users should expect. IBM’s recommendation to pair the models with Granite Guardian for production risk detection is guidance from the company, not evidence of measured benefits in a deployment.
What Users Can Check Next
IBM directs prospective users to the models’ model cards for more information on evaluation results, architecture and training. Teams can compare those details with their own requirements, then test the models using representative audio, including the languages, recording conditions and terminology they expect to process. That testing is needed to determine whether public benchmark results carry over to a particular deployment.
The next useful evidence will include fuller benchmark details, clearly stated evaluation conditions and independent or customer testing across varied speech settings. Until then, IBM’s figures describe results the company reports on named public tests; real-world performance remains dependent on the audio and task.
Key Questions
What did IBM announce for Granite Speech?
IBM announced one new model and revisions to two existing models: a multilingual 2-billion-parameter model, an updated Turbo CTC model for transcription, and revised weights for a model focused on 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, and translation from English into Italian and Mandarin.
How much faster is the updated Turbo CTC model?
IBM reports 19,870 RTFx on the Open ASR Leaderboard, compared with about 12,600 at the model’s August launch. IBM says accuracy remained at the same level. The supplied information does not specify the hardware and deployment conditions behind the speed result.
Are the reported performance gains independently confirmed?
Not in the supplied material. The figures are reported by IBM and include comparisons on public test sets. They do not establish performance for every accent, recording environment or business use case.
What license and software support do the models have?
IBM says all three models are released under the Apache 2.0 license and have native support in Transformers. The 5.0 2B model is also listed as supported in vLLM, llama.cpp and mlx-audio for Apple Silicon.
Primary source: Hugging Face · via ThorstenMeyerAI.com
![Claude Code Reads AGENTS.md Only When Telemetry Is On [Fixed]](https://avaoroi.com/wp-content/uploads/2026/09/claude-code-reads-agents-md-only-when-telemetry-is-on-fixed-featured-380x200.jpg)
