IBM Releases Granite Speech 5.0 With Faster Multilingual Models

IBM Granite Speech 5.0 release

IBM just dropped three updates to its Granite Speech lineup, handing developers a fresh 2-billion-parameter speech-language model plus refreshed versions of two workhorses that already see heavy community use. The news landed on Hugging Face on October 21, 2026, and it covers the whole gamut of enterprise speech workloads: high-accuracy transcription and translation, ultra-fast on-device inference, and speaker-attributed records for meetings and calls. All three ship under the Apache 2.0 license and work natively in Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.

Granite Speech 5.0 2B model details

The headliner is **granite-speech-5.0-2b**, a direct successor to the 4.1 2B model at the same parameter count. It transcribes English, French, German, Spanish, Portuguese, and Japanese, and translates to and from English in those languages plus English to Italian and Mandarin. Training leaned on 390,000 hours of public audio alongside synthetic data crafted for Japanese ASR, keyword biasing, and translation tasks. Compared with its predecessor, the new model delivers higher accuracy on English short-form benchmarks and handles noise, reverberation, and distant-microphone speech more gracefully. Developers can now steer multilingual transcription by simply naming the input language in the prompt, and the model sustains translation quality over longer, more conversational passages rather than just short read sentences. Keyword biasing supports lists of up to 2,000 terms across every supported language, and punctuation with capitalization can be requested in the prompt-capitalization F1 scores range from 89.7 to 99.5 across internal test sets. On the Open ASR Leaderboard’s public English short-form test sets, the model averages a word error rate of 5.06 percent, placing it second among open-weight ASR models as of the October 21 evaluation date.

The speed-focused **granite-speech-5.0-470m-turboctc** gets its first major revision since its August debut. Originally an English-only 470-million-parameter encoder built for throughput, the updated version adds multilingual transcription for English, Japanese, Spanish, German, French, Italian, and Portuguese. Transcripts now emerge with native punctuation and capitalization, removing the need for a separate truecasing step. Training that explicitly targets reverberant audio makes the model more robust in echoey rooms and far-field recordings. Perhaps most striking is the throughput gain: by compiling a single Conformer block and reusing its graph across the stack, IBM cut warmup time and pushed the model to 19,870 real-time factor (RTFx) on the Open ASR Leaderboard, up from roughly 12,600 RTFx at launch-a 1.6× speedup at the same accuracy level. The company also provides a WebGPU streaming demo so developers can try the model directly in the browser.

The third update, **granite-speech-4.1-2b-plus**, keeps the 4.1 architecture, prompt format, and API unchanged so it serves as a drop-in replacement with improved weights. It targets the recordings enterprises actually wrestle with: meetings, support calls, and interviews with multiple speakers running long. Speaker attribution improves noticeably-word diarization error rate falls 7 percent and speaker-attributed word error rate drops 7 percent overall. Timestamp precision sees the largest leap: average timestamp shift across all speech is cut by more than half, from 210.86 to 98.54, while short-speech shift edges down from 29.27 to 26.20. Long-form word error rate declines 24 percent, from 13.53 to 10.22, and keyword-biasing errors on 20-term lists fall 45 percent, from 11.79 to 6.48. These gains make the model a stronger fit for captioning, searchable archives, and compliance review where timing and speaker identity matter as much as raw transcription accuracy.

Choose Granite Speech model guide

For teams deciding which model to adopt, IBM offers a straightforward rubric. Choose **granite-speech-5.0-2b** when transcription accuracy, translation, or large keyword lists are the priority. Pick **granite-speech-5.0-470m-turboctc** for raw speed or on-device deployment. Reach for **granite-speech-4.1-2b-plus** when working with multi-speaker recordings such as meetings or call-center logs. Each model card publishes full evaluation results, architecture details, training data descriptions, and usage examples. IBM also recommends pairing the speech models with Granite Guardian for production deployments that need additional risk detection around toxicity and hallucinations.

Open weight speech models enterprise

The releases arrive at a moment when open-weight speech models are closing the gap with proprietary APIs on both accuracy and latency, while offering the licensing freedom and deployment flexibility that enterprise teams require. By delivering a high-accuracy multilingual model, a multilingual speed champion, and a speaker-aware specialist in a single coordinated update, IBM gives builders a complete toolkit that can be mixed and matched across a product line without juggling incompatible APIs or licenses. Developers can pull the models today from the ibm-granite organization on Hugging Face and integrate them with the inference stack of their choice.

Frequently asked questions

What are the three new Granite Speech models IBM released on October 21, 2026?

IBM released granite-speech-5.0-2b for high-accuracy multilingual transcription and translation, granite-speech-5.0-470m-turboctc for ultra-fast on-device inference, and granite-speech-4.1-2b-plus as a drop-in replacement with improved speaker attribution for meetings and calls.

Which languages does the granite-speech-5.0-2b model support for transcription and translation?

The model transcribes English, French, German, Spanish, Portuguese, and Japanese. It translates to and from English in those six languages plus English to Italian and Mandarin.

How much faster is the updated granite-speech-5.0-470m-turboctc compared to its August launch version?

The updated model achieves 19,870 real-time factor (RTFx) on the Open ASR Leaderboard, a 1.6× speedup from the roughly 12,600 RTFx at launch, while maintaining the same accuracy level.

What improvements does granite-speech-4.1-2b-plus bring for multi-speaker recordings?

It reduces word diarization error rate by 7 percent, speaker-attributed word error rate by 7 percent, cuts average timestamp shift by more than half (210.86 to 98.54), lowers long-form word error rate by 24 percent (13.53 to 10.22), and drops keyword-biasing errors on 20-term lists by 45 percent.

Where can developers download these models and what license do they use?

All three models are available from the ibm-granite organization on Hugging Face under the Apache 2.0 license and work natively with Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.

Source: Hugging Face