IBM releases Granite Speech 5.0 for open multilingual transcription

IBM released Granite Speech 5.0 on October 21, expanding its downloadable speech-model family across three different tasks rather than replacing every model with a single checkpoint. The centerpiece is Granite Speech 5.0 2B, an Apache-2.0 speech-language model designed for multilingual automatic speech recognition and speech translation. Two updated companion models address low-latency transcription and speaker-labelled recordings. This separation matters because an application optimized for meeting notes has different constraints from an edge transcription service.

The new 2B checkpoint transcribes English, French, German, Spanish, Portuguese and Japanese, and supports translation to and from English across those languages. IBM also describes English-to-Italian and English-to-Mandarin translation. It reports training on 390,000 hours of publicly sourced audio plus task-specific synthetic examples. Language-ID prompts, configurable punctuation and terminology lists of up to 2,000 entries are intended to help with accents, product names and specialist vocabulary.

For lighter deployments, the updated Granite Speech 5.0 470M Turbo CTC adds multilingual recognition, punctuation and improved robustness to reverberation. The updated Granite Speech 4.1 2B Plus instead focuses on who spoke and when, including speaker attribution and tighter timestamps. IBM reports a Turbo CTC speed figure of 19,870 times real time on a specified benchmark, but throughput depends strongly on hardware and testing setup. Those claims are not independent Open Model Weights measurements.

Our assessment: this is a useful open-weight toolkit for organizations that want their own speech processing rather than opaque hosted transcription. Before deployment, evaluate language-specific word error rate, consent and retention policies for recorded voices, timestamps in overlapping conversations and real runtime memory. Apache-2.0 distribution is a meaningful advantage, but source data coverage and performance on an organization's actual microphones still require verification.