IBM's 1B-Parameter Speech Model Tops OpenASR Leaderboard

March 11, 2026

IBM's 1B-Parameter Speech Model Tops OpenASR Leaderboard

Published: March 11, 2026 at 12:30 AM

Updated: March 11, 2026 at 12:30 AM

100-word summary

IBM released Granite 4.0 1B Speech, a model half the size of its predecessor that ranks first on the OpenASR leaderboard. The 1-billion-parameter model transcribes speech in six languages and translates to and from English, all while running offline on your device. No cloud needed. The surprising part: shrinking the model actually improved English transcription accuracy. It now handles Japanese alongside English, French, German, Spanish, and Portuguese, with faster inference through speculative decoding. For companies nervous about sending voice data to the cloud, this opens the door to meeting transcription and field recording without the privacy trade-off. Apache 2.0 license means you can ship it in your product tomorrow.

What happened

IBM released Granite 4.0 1B Speech, a model half the size of its predecessor that ranks first on the OpenASR leaderboard. The 1-billion-parameter model transcribes speech in six languages and translates to and from English, all while running offline on your device. No cloud needed. The surprising part: shrinking the model actually improved English transcription accuracy. It now handles Japanese alongside English, French, German, Spanish, and Portuguese, with faster inference through speculative decoding.

Why it matters

For companies nervous about sending voice data to the cloud, this opens the door to meeting transcription and field recording without the privacy trade-off. Apache 2.0 license means you can ship it in your product tomorrow.

Sources