IBM AI Releases Granite Speech 5.0 2B for Multilingual Transcription and Translation
MarkTechPost Asif Razzaq ● Covered by 2 sources
IBM released Granite Speech 5.0 2B, an open model that transcribes and translates speech in 6 languages. It’s Apache 2.0 and can run on GPU, Mac, or even local CPU.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
IBM’s new Granite Speech 5.0 2B is an open speech model built for transcription and translation, and the pitch is broader than “just another ASR model.” It handles 6 languages, ships under Apache 2.0, and IBM is positioning it for accurate transcription, fast on-device use, and speaker-attributed output.
The model itself is compact by current standards at 2 billion parameters, but it is not a toy. IBM says the LLM backbone was extended to 16k positions with YaRN, and the company recommends audio chunks of about 2 minutes. Developers can run it on GPUs through Transformers or vLLM, on Apple Silicon with mlx-audio, or locally on CPU and Mac through llama.cpp, with Q8_0 recommended for GGUF.
IBM also changed two existing models. Turbo CTC keeps its 470M size but now transcribes 7 languages with native punctuation and capitalization, while IBM says throughput rose to 19,870 RTFx from about 12,600 at launch. The updated Granite Speech 4.1 2B Plus is a drop-in weight update too, with average timestamp shift falling from 210.86 to 98.54 and long-form WER improving from 13.53 to 10.22.
On the benchmark side, IBM’s own chart puts Granite Speech 5.0 2B at 5.68% average English WER on the Open ASR Leaderboard, which makes it the #2 open-weight model there behind Qwen3-ASR-1.7B at 5.40%. IBM’s blog also cites 5.06%, but that number is unofficial and based only on public short-form sets. The spread across test sets is large: 0.99% on LibriSpeech Clean, 2.21% on LibriSpeech Other, and 13.79% on the private conversational set.
The model seems strongest when it can use its extra controls. IBM says keyword biasing can take lists of up to 2,000 terms, and on Earnings-22 that lifts keyword F1 from 67.2 to 90.1 with correct terms. Japanese is the weakest language in the multilingual set at 10.74% WER, while Portuguese and Spanish land at 2.60% and 3.50% on CommonVoice.
My take — AI-written commentary, not fact-checked reporting
This is the useful kind of AI release: smaller, open, deployable, and not pretending every speech problem is the same one. The bigger story is that keyword biasing and multilingual support matter more to real users than another leaderboard victory lap. Fancy demos are cheap; getting product names, acronyms, and noisy audio right is the actual job.
Read more about this at: MarkTechPost
