IBM AI Releases Granite Speech 5.0 2B for Multilingual Transcription and Translation

IBM released Granite Speech 5.0 2B, a two-billion-parameter open model for transcription and translation across six languages, under the Apache 2.0 license. It runs on GPUs through Transformers or vLLM, on Apple Silicon, and on a local CPU through llama.cpp. IBM also refreshed two companion models: a fast edge-focused transcriber and a version built for telling speakers apart.

Speech-to-text is one of the least glamorous and most widely deployed parts of AI. Call centers, courts, clinics, newsrooms and accessibility tools all lean on it. A capable model that can run on hardware an organization already owns, without sending audio to someone else's cloud, is a meaningful option.

We like this release for what it is: practical. On the Open ASR Leaderboard it sits second among open-weight models with a 5.68% average English word error rate, behind Qwen3-ASR-1.7B at 5.40%. That's not a crown, and IBM's own blog cited a rosier 5.06% figure from an unofficial, narrower test set. We'd go with the leaderboard number and treat the blog figure as marketing.

The feature I'd watch most closely is keyword biasing. Organizations can feed the model lists of up to 2,000 terms, such as product names, people and acronyms. On an earnings-call benchmark, keyword accuracy jumped from 67.2 to 90.1 F1 with correct lists, and still reached 82.8 when the lists mixed right and wrong terms. That is exactly the failure mode that makes transcripts useless in specialized work: the model gets the common words and mangles the ones that matter.

Now the honest limits. Conversational audio is still hard. On the leaderboard's private conversational set, error rates climb to 13.79%, and Japanese is the weakest language at 10.74% on CommonVoice. IBM recommends segments of about two minutes, and warns that lower-bit quantized versions lose accuracy on audio longer than about a minute. Anyone running this on cheap hardware to save money should test on their own recordings before trusting it.

There's also the bigger picture of what cheap, local transcription enables. Accessibility and record-keeping get better. So does the capacity to quietly transcribe every meeting, call and conversation in a workplace. The technology is neutral; the consent practices around it are not. As speech models get small enough to run anywhere, I'm watching whether the norms for recording people keep pace.

Our take: a solid, openly licensed workhorse, especially for teams that need on-premises control and domain vocabulary. Not the accuracy leader, and not a reason to skip your own evaluation.

Per the model card, Granite Speech 5.0 2B pairs a Conformer speech encoder with a language model derived from IBM's granite-4.0-1b-base, extended to 16k positions. It was trained on 390,000 hours of public audio plus synthetic data over 40 days on eight H100 GPUs. The updated Turbo CTC model, at 470M parameters, now covers seven languages and reached a throughput of 19,870 RTFx, while the refreshed 4.1 2B Plus cut its average timestamp shift roughly in half.