IBM has released Granite Speech 5.0 2B, a new open-weight speech-language model that converts speech into text and translates across six languages. The 2-billion-parameter model ships under the Apache 2.0 license and runs on GPUs, Apple Silicon, and even local CPUs, making it one of the most accessible high-accuracy speech models available today.
The release lands strong on public benchmarks. Granite Speech 5.0 2B ranks as the number-two open-weight model on the Open ASR Leaderboard with a 5.68% average English word error rate, trailing only Qwen3-ASR-1.7B. It scores as low as 0.99% on the LibriSpeech Clean test set, though accuracy drops on conversational audio, with 13.79% on the leaderboard’s private conversational set.
Beyond raw accuracy, the update adds practical features: prompting the model with the input language, keyword biasing with lists of up to 2,000 terms for product names and acronyms, on-demand punctuation and capitalization, and improved robustness to noise and distant microphones. The model was trained on 390,000 hours of public audio plus synthetic data, using eight H100 GPUs over 40 days.
IBM also refreshed two existing models. The edge-focused Turbo CTC keeps its 470-million-parameter footprint while jumping to 19,870 RTFx throughput and adding a seventh language, and a weight update to Granite Speech 4.1 2B Plus sharply improves timestamp accuracy and long-form transcription.
For developers comparing speech-to-text options, the open-weight release is a strong alternative to commercial APIs like ElevenLabs, AssemblyAI, and Descript, especially for on-device or privacy-sensitive deployments. The model is available now on Hugging Face.
Source: Marktechpost
