TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
IBM announced Granite Speech model updates: a new 2-billion-parameter model for transcription and translation, a faster multilingual Turbo CTC model, and revised weights for a model focused on multi-speaker recordings. IBM reports gains on its cited benchmarks, but the results are vendor-reported and do not establish performance across every real-world deployment.
IBM has announced one new Granite Speech model and updates to two existing models, adding capabilities for multilingual transcription and translation, faster speech recognition, and improved speaker labeling. The release matters to organizations handling audio because the three models target different workloads, from high-throughput transcription to long recordings where identifying who spoke and when is important.
The new granite-speech-5.0-2b is a 2-billion-parameter successor to granite-speech-4.1-2b. 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 to Italian and Mandarin. The company reports training it on 390,000 hours of public audio, alongside synthetic data for Japanese speech recognition, keyword biasing and translation.
IBM says the model improves English short-form recognition and handling of noise, reverberation and distant microphones compared with its predecessor. It also supports language identification through prompts and keyword lists of up to 2,000 terms, which can steer recognition toward names, product terms or acronyms. The company reports capitalization F1 scores from 89.7 to 99.5 across its test sets when formatted output is requested. 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 retains its 470-million-parameter footprint and adds transcription in English, Japanese, Spanish, German, French, Italian and Portuguese, as well as native punctuation and capitalization. IBM says it reached 19,870 RTFx on the Open ASR Leaderboard, compared with roughly 12,600 at its August launch, at the same accuracy level. RTFx measures processing speed relative to audio duration; IBM’s comparison covers English short-form speech across nine public test sets.
The third model, granite-speech-4.1-2b-plus, has revised weights for meetings, calls and other recordings with multiple speakers. IBM reports a 7% reduction in both word diarization error rate and speaker-attributed word error rate. It also reports a 24% reduction in word error rate on long-form test sets, and a 45% reduction in errors on biased keywords when using 20-term lists. The company says the model retains the same architecture, prompt format and API, making it a drop-in replacement for the earlier version.
Three Models for Different Audio Workloads
The release gives developers and organizations three distinct options rather than a single model intended for every audio task. IBM recommends the new 2B model for accuracy, translation and keyword biasing; Turbo CTC for speed or on-device use; and the revised 4.1 2B Plus for recordings with several speakers. That separation can help teams choose based on whether latency, language coverage or speaker attribution is their main requirement.
Better speaker labels and timestamps may make recorded meetings and calls easier to search, caption and review. Improved recognition of specialized terms could also reduce correction work in audio containing product names or acronyms. Those are practical potential benefits, not guarantees: performance will depend on the recording conditions, language, vocabulary and evaluation setup.
For developers, the Apache 2.0 license and support across several inference frameworks allow use in a range of deployment environments. IBM also points to a WebGPU streaming demo for Turbo CTC and recommends pairing the models with Granite Guardian in production deployments that need additional risk detection. The announcement does not provide comparative operating costs or independent testing of that recommendation.
As an affiliate, we earn on qualifying purchases.
How the Granite Releases Differ
The announcement builds on IBM’s existing Granite Speech collection. The new 2B model succeeds granite-speech-4.1-2b, while Turbo CTC is an updated version of a 470-million-parameter model IBM introduced in August as an English-only, speed-focused system. The revised 4.1 2B Plus keeps its version name because IBM says its architecture, API and prompt format have not changed.
IBM’s report includes leaderboard and model-card metrics, but the cited figures should be read within their stated scope. For example, the 5.06% average English word error rate cited for granite-speech-5.0-2b is an average across public English short-form test sets on the Open ASR Leaderboard, with results dated October 21, 2026. IBM says that places the model second among top open-weight ASR models in that comparison. It is not a measure of performance across all languages, audio types or production uses.
The model cards are identified by IBM as the source for full evaluation results, training data, architecture details and usage examples. The announcement supplies selected metrics and describes them as company evaluation results; it does not include independent validation in the source material.
“Together they cover the full range of enterprise speech work.”
— IBM, in its Granite Speech announcement
multilingual speech recognition device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Results and Deployment Limits
The announcement reports benchmark gains, but the source material does not provide independent verification of those results. It also does not establish how the models compare across all supported languages, accents, noisy environments or specialized domains. The cited leaderboard figures cover specified English short-form test sets and should not be treated as universal performance measures.
IBM does not state in the announcement how much computing capacity the models require in different deployment settings, what their operating costs are, or how results vary on customer recordings. Those questions may affect whether the speed and accuracy claims translate into benefits for a particular organization. The precise evaluation methods and broader results are directed to the model cards.
As an affiliate, we earn on qualifying purchases.
Model Cards and Testing Ahead
Developers can review the models’ model cards and usage examples for fuller evaluation results, architecture information, training data and setup guidance. IBM also offers a browser-based WebGPU streaming demo for Turbo CTC. The company has invited users to try the releases and provide feedback.
The next useful evidence will come from testing on recordings that reflect each organization’s languages, acoustics and vocabulary, alongside further disclosed evaluation details. The announcement does not specify a later release date or another scheduled milestone.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did IBM announce?
IBM announced granite-speech-5.0-2b, a new speech model, and updates to granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus.
Which model is aimed at translation?
IBM positions granite-speech-5.0-2b for multilingual transcription and translation, including translation to and from English in several supported languages.
What changed in Turbo CTC?
The 470-million-parameter model adds multilingual transcription and native punctuation and capitalization. IBM reports 19,870 RTFx on the cited leaderboard, up from roughly 12,600 at its August launch.
Which model is intended for meetings with multiple speakers?
IBM recommends the revised granite-speech-4.1-2b-plus for meetings, calls and other multi-speaker recordings, citing improvements to speaker attribution, timestamps and long-speech performance.
Are the reported benchmark results independently verified?
The supplied announcement presents IBM-reported results and points readers to model cards for more detail. It does not provide independent verification of the figures.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
