IBM releases Granite Speech 5.0 2B for multilingual transcription and translation
IBM’s Apache-2.0 speech model supports six-language transcription and multilingual translation, with vendor-reported Open ASR results and local-serving requirements.
2 min read
IBM has released Granite Speech 5.0 2B, a downloadable speech model for multilingual transcription and translation. The model transcribes English, French, German, Spanish, Portuguese and Japanese. It can translate those languages into English, translate English into them, and also translate English into Italian and Mandarin.
IBM says it trained the 2-billion-parameter model on 390,000 hours of public and synthetically generated audio. New controls include language-ID prompting, formatting instructions and a keyword list of up to 2,000 terms for domain vocabulary. The model works with Transformers and IBM’s helper library, while IBM also lists vLLM, llama.cpp and mlx-audio integrations.
The weights are available under Apache 2.0. Hugging Face showed no hosted inference provider when checked on October 22, so users need to run the model locally or through another serving stack. IBM recommends splitting audio into segments of roughly two minutes or less.
IBM reports an average English word-error rate of 5.06 across the public short-form datasets in the Open ASR Leaderboard, placing the model second among the open-weight systems it compared. That is a vendor-reported result and should not be read as proof of equivalent performance on the other supported languages, long recordings or noisy production audio.
The canonical helper repository is maintained by IBM Granite under Apache 2.0. It had 2 GitHub stars and 4 forks on October 22; those are total counts, not evidence of measured growth.
Image credit: Photo by AltumCode on Unsplash, used under the Unsplash License.
Sources: IBM announcement · Model card · Canonical repository · Image · Unsplash License
Conversation
0 approved commentsNo comments yet. Start a thoughtful conversation.
Join the conversation.
Sign in with ChatGPT to leave a comment or rate this article.