IBM’s model card dates the release of Granite Speech 5.0 2B to October 20, 2026. The dense 2-billion-parameter model is licensed under Apache 2.0 and handles automatic speech recognition (ASR), which turns speech into text, and speech translation. IBM lists six transcription languages and specific translation directions, plus a hosted endpoint for trying the model without installing its weights locally.
What Granite Speech 5.0 2B can transcribe and translate
IBM lists English, French, German, Spanish, Portuguese and Japanese for ASR. For translation, it documents speech to and from English for French, German, Spanish, Portuguese and Japanese, as well as English-to-Italian and English-to-Mandarin.
| Documented direction | Languages |
| English to and from | French, German, Spanish, Portuguese, Japanese |
| English to | Italian |
| English to | Mandarin |
The listed directions matter when choosing a model for a specific language pair: IBM names English-to-Italian and English-to-Mandarin, but does not list the reverse directions. The model card also documents prompt-based keyword biasing for lists of up to 2,000 terms, which can help transcripts handle specialized vocabulary.
How IBM built the model, and its audio guidance
IBM describes a 16-block Conformer encoder trained with CTC, followed by a two-layer window q-former projector. The pipeline downsamples the audio representation by 10× and supplies 10 Hz acoustic embeddings to the language model.
For best results, IBM recommends splitting longer recordings into audio segments of about two minutes or less. Its separate GGUF guidance recommends the Q8_0 quantization; IBM says lower-bit quantizations lose accuracy on audio longer than about one minute. The sample Transformers code expects mono, 16 kHz audio.
What the reported WER figures measure
Word error rate (WER) measures transcription errors as a percentage, with lower scores indicating fewer errors in the evaluated material. MarkTechPost reported a 5.68% average English WER for Granite Speech 5.0 2B on the Open ASR Leaderboard as of October 21, 2026, across nine public and two private test sets.
The same publication reported a separate 5.06% figure for public short-form sets, dated October 15, 2026. The figures refer to different evaluation dates and test-set scopes. MarkTechPost also reported a 14.80% average WER on the FFASR leaderboard as of October 18, 2026, where Granite ranked fourth among 42 models.
How to try and run it
IBM documents a hosted Hugging Face Inference Endpoint with an OpenAI-compatible API, so developers can try requests without first installing the model locally. The endpoint scales to zero when idle, and IBM says the first request may take several minutes to start.
For local use, IBM lists Transformers, vLLM, llama.cpp with GGUF, and mlx-audio for Apple Silicon. These give developers several documented runtime paths; the sample Transformers workflow expects mono 16 kHz audio, and IBM’s segment and quantization guidance applies when preparing longer recordings or choosing a GGUF build.
Ray Codes’ screencast walks through a local Python setup and sample transcription and French-to-English translation.
IBM’s prompt and output guidance
IBM says unfamiliar or malformed prompts fall back to transcription and recommends the model for speech recognition and translation. The company also recommends pairing it with Granite Guardian to detect and flag risks in prompts and responses, and cautions that outputs may be inaccurate, biased, offensive or unwanted.
