IBM Launches Granite Speech 5.0: 5.06% WER, 19,870 RTFx Speed

Executive Overview

IBM has released three updates to its Granite Speech family via the official Hugging Face blog: the new granite-speech-5.0-2b with a 5.06% average WER, the granite-speech-5.0-470m-turboctc reaching 19,870 RTFx, and the granite-speech-4.1-2b-plus cutting long-form transcription errors by 24%. All three models are immediately available under the Apache 2.0 open-source license on Hugging Face, with native support for Transformers, vLLM, llama.cpp, and mlx-audio. The release positions IBM's open-weight ASR stack as a direct competitor to proprietary speech APIs, though Arabic remains absent from the officially supported language list.

Deep-Dive Features & Architecture

granite-speech-5.0-2b succeeds granite-speech-4.1-2b with the same 2B parameter count and was trained on 390,000 hours of public audio plus synthetic data tailored for Japanese speech recognition, keyword biasing, and translation. Key upgrades include:

• Higher accuracy on short-form English ASR
• Improved robustness to noise, echo, and far-field microphone distances
• Input language steering via prompt
• Keyword lists of up to 2,000 terms per supported language
• Automatic punctuation and capitalization on request with F1 scores between 89.7 and 99.5

granite-speech-5.0-470m-turboctc retains a 470M parameter footprint while adding multilingual support for seven languages, native punctuation and capitalization without a separate processing step, and higher resistance to echo and difficult acoustic environments. Speed jumped to 19,870 RTFx thanks to fusing a single Conformer block and reusing its graph across layers, which reduced warm-up time and raised throughput 1.6x at the same accuracy.

granite-speech-4.1-2b-plus targets enterprise meetings, calls, and multi-speaker interviews while preserving the same architecture, prompt format, and API as a drop-in replacement for the previous model. It recorded a 7% improvement in WDER and 7% in cpWER, cut AAS from 210.86 to 98.54 (more than half), reduced long-form WER by 24% from 13.53 to 10.22, and cut keyword errors by 45% from 11.79 to 6.48.

Benchmark & Competitive Performance

On the Open ASR Leaderboard (nine general short-form English test sets, October 2026 data), granite-speech-5.0-2b ranked second among the best open-weight ASR models with an average WER of 5.06%, surpassing its 4.1-2b predecessor. On the speed axis, turboctc outperformed with 19,870 RTFx versus roughly 12,600 RTFx at its August launch—a 1.6x improvement. Meanwhile, 4.1-2b-plus outperformed its previous version across six error metrics simultaneously, most notably cutting AAS by over 53% and improving long-form accuracy by 24%.

Industry Impact & Enterprise Adoption

The release strengthens IBM's position in the open-weight speech recognition market, offering enterprises a free, Apache 2.0 alternative to proprietary ASR APIs from cloud providers. The drop-in compatibility of 4.1-2b-plus is particularly significant for organizations already running Granite Speech in production, as it requires no prompt or API changes. The turboctc model's 19,870 RTFx makes it viable for high-throughput batch transcription and real-time applications, while the experimental WebGPU demo signals IBM's interest in browser-based inference. However, the absence of Arabic support limits adoption in Middle Eastern markets, where demand for open-source ASR is growing.

Conclusion

IBM's Granite Speech 5.0 release delivers measurable gains across accuracy, speed, and long-form robustness—all under a permissive Apache 2.0 license. With 5.06% WER, 19,870 RTFx, and a 24% reduction in long-form errors, the models are competitive with leading open-weight ASR systems. Enterprises seeking a free, self-hostable speech stack should evaluate these models, though Arabic-language deployments will require alternative solutions until IBM expands language coverage.

Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis