🔍 Read the full analysis: A Closer Look At Granite Speech 5.0 For Faster, More Accurate Audio Tasks on ThorstenMeyerAI.com
TL;DR
IBM announced one new model and revisions to two existing models in its Granite Speech collection. The releases target multilingual transcription and translation, faster transcription, and speaker attribution and timestamps; the performance figures cited are IBM’s own evaluations, not independent validation.
IBM announced three Granite Speech model updates for organizations processing audio: a new multilingual 2-billion-parameter model for transcription and translation, a faster multilingual transcription model, and revised weights aimed at improving speaker attribution and timestamps, as detailed in the original analysis. The company reports benchmark gains for the releases, but the figures come from IBM-cited evaluations and do not establish results for every language, recording condition or business workflow.
The new granite-speech-5.0-2b replaces the 4.1 2B model at the same parameter count. IBM says it transcribes English, French, German, Spanish, Portuguese and Japanese, and translates between English and those languages. It also supports translation from English to Italian and Mandarin. According to the company, training used 390,000 hours of public audio as well as 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 supports input-language prompts 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, with support 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 score 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. IBM says that update leaves the architecture, prompt format and API unchanged.
Three Models, Three Audio Priorities
The releases distinguish between workloads that can require different trade-offs. The 5.0 2B model is positioned for multilingual recognition and translation; Turbo CTC targets high-speed transcription; and the 4.1 2B Plus revision focuses on speaker labels and timing in longer or multi-speaker recordings. That gives teams options rather than requiring one model to serve every audio task.
Those features can matter after speech is converted to text. More precise timestamps may help align captions or locate sections in recordings, while speaker labels can make meetings and support calls easier to review. Keyword lists may help with names, acronyms and product terminology. These are plausible uses of the announced features, not results demonstrated in the supplied material for specific customer deployments.
IBM’s figures offer a starting point for comparing the models, but they do not settle how a system will perform on a company’s own audio. Organizations still need to test representative recordings, including the accents, vocabulary, background noise and overlapping speech they encounter, and weigh output quality against review and deployment costs.
multilingual speech recognition software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the Granite Updates Differ
IBM introduced Turbo CTC in August as a 470-million-parameter, English-only model built for speed. The October revision keeps that model size and adds six transcription languages, along with punctuation and capitalization. The new 5.0 2B model takes the multilingual transcription and translation route, while the revised 4.1 2B Plus is aimed at diarization, timing and longer recordings.
IBM says the 5.0 2B model ranked second among open-weight ASR models on average English word error rate, at 5.06%. The company labels this result unofficial and says it averages public English short-form test sets on the Open ASR Leaderboard as of October 15, 2026. Its Turbo CTC speed comparison uses nine public English short-form test sets and is also dated to October 2026. Lower word error rate and higher RTFx indicate better performance on those respective measures, but they capture different aspects of a model.
All three models are released under Apache 2.0 and have native support in Transformers, according to IBM. The company directs users to model cards for evaluation, architecture and training details, and recommends pairing the models with Granite Guardian for production deployments that require 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
audio transcription software with speaker labeling
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmark Limits and Open Questions
The announcement’s figures have not been independently validated in the supplied material. Public benchmark results, especially those based on short-form English test sets, do not show how the models perform across all accents, language varieties, specialized vocabulary, noisy environments or overlapping speech. IBM’s reported 5.0 2B capitalization F1 scores are presented as a range across test sets, without a breakdown here.
IBM describes Turbo CTC’s accuracy as unchanged while reporting a higher RTFx score, but the source material does not include the full results table needed to examine that comparison. It also does not specify the hardware or deployment conditions behind the speed figures. The long-form improvements for the 4.1 2B Plus update are company-reported; details of the test sets and the underlying results are not included in the material provided.
It remains unclear how much the reported gains will carry over to individual organizations’ audio and workflows. The announcement also does not quantify the effect of keyword biasing, improved timestamps or speaker labeling in customer deployments.
automatic speech to text transcription device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing Before Production Use
IBM points prospective users to the models’ model cards for evaluation results, architecture and training information. Organizations considering deployment can compare the three models against their own recordings and requirements, paying particular attention to language coverage, noise, multiple speakers, specialist terms and the need for human review.
No further release milestone or independent evaluation is specified in the supplied material. The next evidence that would clarify the announcement’s practical reach is more detailed benchmark documentation and testing on audio beyond the public sets cited by IBM.
professional audio transcription tools
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 Granite Speech model and revised versions of two existing models. They target 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 between English and those languages, plus translation from English to Italian and Mandarin.
How much faster is the Turbo CTC model, according to IBM?
IBM reports 19,870 RTFx on the Open ASR Leaderboard, compared with about 12,600 at the model’s August launch. The announcement says accuracy was maintained, but the figures are IBM-reported and the supplied material does not detail the hardware conditions.
Are the reported improvements independently confirmed?
The source material does not provide independent validation. IBM’s figures are based on evaluations and public benchmarks it cites, so prospective users should test the models on audio representative of their own needs.
What should organizations test before using the models?
Teams should check performance on their own languages, accents, recording conditions, specialist terms and multi-speaker audio. They should also assess timestamp and speaker-label quality and decide how much human transcript review their workflow requires.
Primary source: Hugging Face · via ThorstenMeyerAI.com
