IBM Releases Granite Speech 5.0 2B AI Model
The new 2-billion parameter model offers multilingual transcription support for developers under an Apache 2.0 license.
Updated on Oct. 22, 2026 in Artificial Intelligence
Live Poll
Do you believe open-source AI models are more beneficial for technological innovation than proprietary ones?
IBM has released its Granite Speech 5.0 2B model, a new speech-to-text tool supporting six languages. This release is now available under an Apache 2.0 license.
Why it matters
The model advances speech processing efficiency by utilizing synthetic data to refine automatic speech recognition and translation capabilities. Its release provides developers with a high-performance open-source option for multilingual transcription tasks.
Granite Speech 5.0 2B features 2 billion parameters and achieved a 0.99% word error rate on the LibriSpeech Clean set. The model was trained using 390,000 hours of audio data over 40 days on eight H100 GPUs.
The players
IBM
A multinational technology corporation focused on enterprise AI, hybrid cloud architecture, and high-performance computing research.
The details
The model architecture employs 16 Conformer blocks—a hybrid of convolutional neural networks and transformers—to process audio. A two-layer window q-former downsamples acoustic embeddings, while LoRA adapters enable fine-tuning of the base granite-4.0-1b-base checkpoint. Training incorporated synthetic datasets to specifically optimize for Japanese speech recognition and keyword biasing, allowing the system to track up to 2,000 keywords.
Timeline
October 15, 2026: IBM conducted internal performance reporting.
October 21, 2026: Leaderboard performance metrics were collected.
October 22, 2026: Granite Speech 5.0 2B was officially released.
The Tech Race
This release positions IBM within the competitive field of compact, open-source speech models that prioritize low error rates. It follows the performance standards set by the Open ASR Leaderboard to validate its efficiency against current industry benchmarks.
Developers and enterprises can immediately integrate the Apache 2.0-licensed model into existing transcription workflows. The system is designed for broad language support, utilizing specialized keyword biasing for improved recognition in professional or technical applications.
The takeaway
IBM provides a new benchmark for lightweight speech models that developers can utilize for multilingual transcription. Monitor the Open ASR Leaderboard for updates to see how subsequent iterations of the Granite series improve upon the current 5.68% average word error rate.
Further reading
For more on the latest research in this field, visit the Artificial Intelligence section.
Live Poll
Do you believe open-source AI models are more beneficial for technological innovation than proprietary ones?
