🔍 Read the full analysis: Granite Speech 5.0: What’s New In Transcription, Translation, And Attribution on ThorstenMeyerAI.com
TL;DR
IBM announced one new Granite Speech model and revised weights for two existing models, targeting multilingual transcription and translation, faster speech recognition, and improved speaker labels and timestamps. Performance figures are IBM-reported benchmark results; the announcement does not establish how the models will perform across every language variety or customer workload.
IBM has announced three updates to its Granite Speech collection: a new multilingual 2-billion-parameter model for transcription and translation, an expanded and faster version of its 470-million-parameter Turbo CTC model, and revised weights for a model focused on speaker attribution and timestamps. The releases target different speech-processing tasks, while the performance improvements described so far are IBM-reported evaluation results, 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 translates to and from English for those languages. It also supports translation from English to Italian and Mandarin. The company says it trained the model on 390,000 hours of public audio, alongside synthetic data for Japanese speech recognition, keyword biasing and translation.
IBM says the 5.0 model is designed to handle noise, reverberation and distant microphones better than its predecessor. Users can specify the input language in prompts and provide keyword lists of up to 2,000 terms to guide recognition and translation. When formatted output is requested, IBM reports capitalization F1 scores ranging from 89.7 to 99.5 across its test sets. It says the model is available under the Apache 2.0 license, with support 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, plus native punctuation and capitalization. IBM reports a speed score of 19,870 RTFx on the Open ASR Leaderboard, compared with about 12,600 at the model’s August launch, while saying accuracy remained at the same level. The updated granite-speech-4.1-2b-plus is aimed at recordings with multiple speakers: IBM reports a 24% reduction in word error rate on long-form test sets, a 7% reduction in two speaker-attribution error measures, and more than a halving of average timestamp shift. The company says its architecture, prompt format and API remain unchanged.
Three Models, Different Audio Needs
The releases reflect that translation, high-volume transcription and multi-speaker recordings place different demands on speech systems. Organizations could select the 5.0 2B model for multilingual recognition and translation, Turbo CTC when throughput is a priority, or the 4.1 2B Plus update when speaker labels and timing are important. These are intended uses based on IBM’s descriptions, not demonstrated outcomes in specific customer deployments.
More accurate timestamps can help align captions or make audio easier to search, while speaker attribution can help people review meeting or support-call records. Keyword biasing may be useful for names, acronyms and product terminology that general-purpose recognition can miss. But benchmark scores alone do not predict results for a particular organization. Teams still need to test the models on representative recordings and account for errors, human review and deployment costs.
IBM reports that the 5.0 2B model averaged 5.06% English word error rate, placing it second among open-weight ASR models in the company’s comparison. IBM labels the result unofficial and says it uses public English short-form test sets on the Open ASR Leaderboard as of October 15, 2026. The company’s reported Turbo CTC RTFx comparison also uses nine public English short-form test sets. Word error rate and RTFx measure different things: lower word error rate indicates fewer recognition errors, while higher RTFx indicates faster processing relative to audio duration.
As an affiliate, we earn on qualifying purchases.
How the Granite Lineup Changed
The three announcements update distinct parts of IBM’s Granite Speech collection, which supports speech recognition and related tasks. IBM introduced Turbo CTC in August as an English-only, 470-million-parameter encoder designed for speed. The October update keeps its parameter count while adding six transcription languages and punctuation and capitalization features.
The new 5.0 2B model replaces the 4.1 2B model as IBM’s multilingual transcription and translation offering, with the same number of parameters. The separate 4.1 2B Plus revision focuses on long recordings, speaker attribution and timestamp accuracy. That division lets users compare models against the job they need to perform rather than treating one system as a fit for every audio workflow.
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 results, architecture and training details, and recommends pairing the models with Granite Guardian in production deployments that need additional risk detection. The supplied announcement does not explain what specific risk-detection results that pairing would provide.
“The model is the successor to granite-speech-4.1-2b at the same parameter count.”
— IBM, in its Granite Speech announcement
multilingual speech recognition tool
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Limits of the Published Evaluations
The supplied material does not include independent validation of IBM’s reported results or enough detail to assess performance across all accents, language varieties, overlapping speech, specialized vocabulary and recording conditions. Public benchmark results can offer a comparison point, but they do not establish performance for every business workflow.
Some parts of the evaluation are not fully detailed in the announcement. IBM gives the 5.0 2B capitalization F1 as a range across test sets without a breakdown here. It says Turbo CTC’s accuracy remained unchanged as its speed score rose, but the supplied material lacks the full benchmark table needed to inspect that comparison. The hardware and deployment conditions behind the RTFx results are also not stated. Customer-level results and adoption details are not provided.
As an affiliate, we earn on qualifying purchases.
What Deployers Should Check
IBM has made model cards and evaluation materials the next point of reference for users who want architecture, training and benchmark details. Organizations considering deployment can compare those materials with their own requirements, then test the models on audio representative of their languages, speakers and environments. That testing can show whether the reported gains carry over to their recordings and whether human correction remains necessary.
The announcement does not give a further release date or a customer deployment milestone. For now, the next evidence will come from fuller evaluation documentation and independent or customer testing. Until those results are available, the announced figures should be read as publisher-reported benchmark claims, not guarantees for real-world transcription quality, translation accuracy or processing speed.
multi-speaker audio transcription software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did IBM announce for Granite Speech?
IBM announced one new model and revisions to two existing models: the multilingual 5.0 2B, the expanded 5.0 470M Turbo CTC, and an update to the 4.1 2B Plus model for 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 to Italian and Mandarin.
What changed in Turbo CTC?
The 470-million-parameter model adds transcription in six languages beyond English, plus native punctuation and capitalization. IBM reports 19,870 RTFx on the Open ASR Leaderboard, compared with about 12,600 at its August launch, and says accuracy was maintained.
Are IBM’s performance results independently verified?
The supplied announcement presents IBM-reported evaluation results and does not provide independent validation. The company’s public benchmark figures also do not establish performance for every accent, recording condition or business use.
What should organizations test before using the models?
Organizations should evaluate the models on recordings similar to their own, including relevant languages, terminology, speaker patterns and acoustic conditions. They should also assess review needs and deployment costs rather than relying on leaderboard figures alone.
Primary source: Hugging Face · via ThorstenMeyerAI.com
