Granite Speech 5.0: transcription guided by names and terminology
IBM’s October 21 release adds language prompting and up to 2,000 keywords for transcription. Russian is absent from the published ASR language list.
The edit is nearly finished, but the subtitles have turned the guest’s surname into another word. The product name needs correcting again. These tiny repairs are exactly where a creator would like to save time. The interesting idea behind Granite Speech 5.0 is giving transcription a list of words that deserve particular attention.
What IBM released
On October 21, 2026, IBM introduced Granite Speech 5.0 2B. It accepts a language hint and up to 2,000 terms to guide transcription and translation. Russian is absent from the published transcription language list. The weights use Apache 2.0. This is a provider announcement, not a VIZURAI integration announcement.
Why a creator might prepare a glossary
Imagine interviewing the founder of a small brand. Their surname, a collection name and a few specialist abbreviations appear throughout the conversation. Those are details the audience and client will notice. Our editorial suggestion is to collect the important spellings before starting the transcript.
Keep that preparation small and specific. A short interview does not need an encyclopedia: identify the words where an error would actually matter. A glossary is still a hint, so reviewing the result remains part of the job.
A small comparison on your own recording
This is a proposed evaluation, not a test we have conducted. Pick a short passage in a supported language containing several unusual words. Save the original audio and the correct spellings, so the reference is clear before you compare outputs.
- Transcribe the passage without a glossary.
- Repeat the same passage with the names and terms supplied.
- Count corrected errors and newly introduced errors separately.
- Listen to uncertain passages before moving the text into your edit.
The useful result is the amount of manual repair still needed on your recording. An impressive demonstration cannot answer that question for you.
What this means for a Russian-language project
An entirely Russian recording is not a reason to switch workflows on the strength of this release. For a foreign-language interview, the glossary idea is worth a small evaluation. We would begin with one complete passage and a straightforward success criterion: names and terms match the original. You still make the final editorial decision.
Create your own result
Choose an available model and check its price in the studio.
Sources
- IBM Granite on Hugging Face — Granite Speech 5.0 Published by the source:

