🔍 Read the full analysis: How Granite Speech 5.0 Makes Transcription, Translation, And Attribution Faster on ThorstenMeyerAI.com
TL;DR
IBM has announced three updates to its Granite Speech collection: a new multilingual 2-billion-parameter model, a faster multilingual 470-million-parameter transcription model, and revised weights for a model focused on long recordings, speaker attribution and timestamps. The performance figures come from IBM’s evaluations and cited public benchmarks; independent validation and results across customers’ real-world audio are not established in the supplied material.
IBM has announced three updates to its Granite Speech collection, adding a multilingual transcription and translation model and revising two others for faster transcription and better speaker attribution and timestamps, as detailed in the original analysis. The releases target distinct speech-processing needs, but performance claims are based on IBM’s evaluations and cited public benchmarks, not independent validation described in the supplied material.
The new granite-speech-5.0-2b has 2 billion parameters, the same size as the 4.1 2B model it succeeds. IBM says it transcribes English, French, German, Spanish, Portuguese and Japanese. It also supports translation to and from English for those languages, plus translation from English to Italian and Mandarin. The company says training used 390,000 hours of public audio, as well as 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 prompt it with the input language and provide 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 revised granite-speech-5.0-470m-turboctc adds transcription in Japanese, Spanish, German, French, Italian and Portuguese alongside English, with native punctuation and capitalization. IBM reports an Open ASR Leaderboard result of 19,870 RTFx, compared with about 12,600 at the model’s August launch, while saying accuracy remained at the same level. RTFx is a speed measure; IBM’s announcement, as described in the source material, does not supply the compute conditions behind that result.
The updated granite-speech-4.1-2b-plus is aimed at long-form and multi-speaker audio. 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 halving average timestamp shift across all speech. The company says the update preserves the model’s architecture, prompt format and API.
Three Models for Different Audio Jobs
The updates reflect that speech systems are often judged on more than whether they can turn audio into words. Organizations may need to translate recordings, process large volumes quickly, or identify who spoke and when. IBM’s lineup separates those goals: 5.0 2B for multilingual transcription and translation, Turbo CTC for speed, and 4.1 2B Plus for speaker and timing information on longer recordings.
Those features could make transcripts easier to search, review or align with captions. Keyword biasing may also help with names, acronyms and product terms. These are potential uses suggested by the announced capabilities, not measured results from customer deployments. The supplied material does not quantify time or cost savings for businesses, or show how much the changes improve any particular workflow.
For buyers, the central question is whether the reported benchmark gains hold on their own recordings. Public test results provide a reference point, but a model’s usefulness may depend on accents, room acoustics, overlapping speech and specialized vocabulary. Teams need to evaluate those factors before treating a speed or accuracy result as evidence of a comparable improvement in production.
professional transcription software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the Granite Releases Differ
IBM’s Granite Speech collection includes models for speech recognition and related audio tasks. The company introduced Turbo CTC in August as an English-only, 470-million-parameter encoder built for speed. The October revision retains that parameter count while adding six transcription languages and punctuation and capitalization features.
The new 5.0 2B model succeeds the 4.1 2B model at the same size and adds the multilingual translation capabilities IBM describes. The revised 4.1 2B Plus, by contrast, focuses on long-form recordings, speaker attribution and timestamp alignment. IBM says all three models use an Apache 2.0 license and have native support in Transformers; it points users to model cards for evaluation results, architecture and training details.
IBM also reports that the 5.0 2B model ranked second among open-weight ASR models on average English word error rate, with a score of 5.06%. The company labels this result unofficial and says it averages public English short-form tests on the Open ASR Leaderboard as of October 15, 2026. Its Turbo CTC speed comparison uses nine public English short-form test sets. Word error rate and RTFx measure different things: the former concerns transcription errors, while the latter represents processing speed.
“The model is the successor to granite-speech-4.1-2b at the same parameter count.”
— IBM, in its Granite Speech announcement
multilingual speech translation device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Questions Beyond IBM’s Benchmarks
The supplied material does not describe independent validation of IBM’s reported results. Public benchmark scores do not establish performance across every accent, language variety, recording condition or business workflow. IBM’s reported leaderboard results cover short-form public test sets, while the 4.1 2B Plus claims concern long-form and speaker-related measures.
Several details needed to assess the comparisons are not included in the source material. These include the full benchmark tables and, for Turbo CTC’s RTFx result, the hardware and deployment conditions used. The 5.0 2B capitalization F1 scores are given as a range across test sets without a breakdown here. It is also unclear how the models perform with overlapping speech, uncommon terminology or noisy audio outside the cited evaluations. The source does not provide customer deployment results or quantified operational savings.
As an affiliate, we earn on qualifying purchases.
What Users Should Check Next
IBM directs users to the models’ model cards for evaluation results, architecture and training details. Teams considering the releases can compare those materials with their own audio and test transcription accuracy, language coverage, processing speed, speaker labels and timestamps under conditions similar to their intended use.
Further evidence would be needed to establish whether the benchmark changes translate into consistent improvements in production. That could include more detailed evaluation tables, deployment conditions and results across varied accents and recording environments. IBM also recommends pairing the models with Granite Guardian for production deployments needing additional risk detection, but the supplied material does not detail the setup or its measured effects.
long-form speech recognition software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did IBM announce?
IBM announced one new model and revisions to two existing Granite Speech models, targeting multilingual speech tasks, faster transcription, and improved speaker attribution and timestamps.
Which languages does Granite Speech 5.0 2B support?
IBM says it transcribes English, French, German, Spanish, Portuguese and Japanese. It supports translation to and from English for those languages, as well as translation from English to Italian and Mandarin.
How much faster is the revised Turbo CTC model?
IBM reports a 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. The supplied material does not state the hardware or deployment conditions for the speed figures.
Are the reported improvements independently verified?
The supplied material reports IBM evaluations and public benchmark figures, but does not describe independent validation. Results may differ on audio with other accents, acoustics, overlapping speakers or specialized vocabulary.
What should teams test before using the models?
Teams should evaluate the models on audio resembling their own work, checking accuracy, language coverage, speed, speaker labels and timestamps. The benchmark results alone do not establish performance or savings in a specific deployment.
Primary source: Hugging Face · via ThorstenMeyerAI.com
