IBM has updated its Granite Speech collection with three open models aimed at different production needs: multilingual transcription and translation, high-speed edge transcription, and long recordings where speaker identity and timing matter.
The new Granite Speech 5.0 2B supports transcription in six languages and translation between English and those languages, plus English output to Italian and Mandarin. It was trained on 390,000 hours of public audio and accepts keyword lists of up to 2,000 terms for names, products and acronyms.
A revised 470-million-parameter Turbo CTC model adds seven-language transcription, punctuation and better handling of echo. IBM reports throughput of 19,870 times real time on the Open ASR Leaderboard, about 1.6 times its launch speed at similar accuracy.
The updated Granite Speech 4.1 2B Plus targets meetings and calls. IBM reports 7% lower speaker-labeling error, more than halved timestamp shift and 24% lower error on long-form tests. All three use Apache 2.0 licensing and support Transformers; the larger 5.0 model also works with vLLM, llama.cpp and Apple Silicon through mlx-audio.