IBM has released updates to its Granite Speech collection, introducing a new SI model and revising two existing systems to improve transcription accuracy, translation capabilities, and speaker attribution for enterprise applications.
What Happened
The release centers on granite-speech-5.0-2b, a new speech-language model designed for high-accuracy multilingual transcription and translation. The model transcribes English, French, German, Spanish, Portuguese, and Japanese, and supports translation to and from English in those languages, as well as English-to-Italian and English-to-Mandarin. IBM states the model was trained on 390,000 hours of public audio and synthetic data. Key new features include language-ID prompting to steer multilingual transcription, support for keyword lists up to 2,000 terms to bias output toward specific names or acronyms, and optional punctuation and capitalization in the output.
IBM also updated granite-speech-5.0-470m-turboctc, an SI tool optimized for speed and on-device use. This revision expands the model from English-only to multilingual support for seven languages and adds native punctuation and capitalization. The company reports that the model now achieves 19,870 RTFx on the Open ASR Leaderboard, a 1.6x speed increase over its launch performance, while maintaining accuracy levels. Additionally, granite-speech-4.1-2b-plus received an update focused on multi-speaker recordings, with improvements in speaker attribution, timestamp accuracy, and long-form speech handling.
Why It Matters
These updates provide developers and enterprises with more precise control over SI systems used for speech processing. The introduction of 2,000-term keyword biasing in granite-speech-5.0-2b addresses a common challenge in automatic speech recognition (ASR) where specific product names or acronyms are misidentified. According to IBM, the new model ranks second among top open-weight ASR models on average English Word Error Rate (WER) at 5.06%, based on official Open ASR Leaderboard results as of October 21, 2026.
For industries relying on meeting transcription and support call analysis, the improvements to granite-speech-4.1-2b-plus offer measurable gains in clarity. IBM reports a 7% reduction in Word Diarization Error Rate (WDER) and a 24% drop in WER on long-form test sets. The update also significantly reduces average timestamp shift, which aids in searching and reviewing audio content. All three models are released under the Apache 2.0 license and support integration with frameworks such as Transformers, vLLM, llama.cpp, and mlx-audio.
The Bottom Line
IBM’s Granite Speech 5.0 updates enhance the accuracy, speed, and control available in open-weight SI models for speech tasks. By improving keyword biasing, multilingual support, and speaker attribution, the releases aim to meet the specific needs of enterprise workflows involving transcription, translation, and multi-speaker analysis.