IBM launches granite speech 5.0 2b for enhanced multilingual transcription and translation

IBM has launched Granite Speech 5.0 2B, a new open-source speech-language model designed for transcription and translation across six languages. The model, released under Apache 2.0, includes updates to two existing models: Turbo CTC and 4.1 2B Plus. It features a speech encoder with 16 Conformer blocks trained using CTC, along with a 2-layer window q-former that downsamples acoustic embeddings by 10x. The LLM is based on granite-4.0-1b-base, extended to 16k positions with YaRN and fine-tuned with LoRA adapters. Training involved 390,000 hours of audio data, achieving a 5.68% average English WER on the Open ASR Leaderboard as of October 21, 2026. The model performs exceptionally well in Portuguese and Spanish, with WERs of 2.60% and 3.50%, respectively, but struggles with Japanese at 10.74%. For translation, X-to-English CoVoST BLEU scores reach 53.6 for Portuguese and 40.1 for German. IBM also improved keyword biasing, increasing keyword F1 scores from 67.2 to 90.1 on Earnings-22. The updated Turbo CTC model now supports seven languages with native punctuation, achieving 19,870 RTFx throughput.