IBM releases Granite Speech 5.0 2B, #2 in its speech test at 5.06% error

IBM's update splits the Granite Speech family into three purpose-built models rather than one general-purpose system. The flagship Granite Speech 5.0 2B keeps the previous model's 2-billion-parameter size but was trained on 390,000 hours of public audio, and IBM is emphasizing robustness to noise and reverberation — conditions that trip up transcription in real meeting rooms and factory floors. A new keyword feature lets users pre-register up to 2,000 terms, aimed at the common failure where names, product names, and acronyms are misheard.

The lightweight Granite Speech 5.0 470M Turbo CTC moves from English-only to seven languages including Japanese, adds automatic punctuation and English capitalization, and runs about 1.6 times faster, which IBM frames around on-device transcription. Separately, the Granite Speech 4.1 2B Plus targets multi-speaker recordings, with speaker-identification error down about 7% and long-audio transcription error down 24% — improvements that matter for subtitling and searching recorded conversations.

Because all three ship under Apache 2.0, companies can deploy and modify them commercially without a proprietary license, which may make them an attractive option for organizations that want speech tools running on their own infrastructure rather than through a cloud API.