IBM’s Granite Speech update separates translation from speaker-labelled transcripts

IBM’s October 21 Granite Speech announcement groups a new speech‑language model with updates to two existing models. The practical choice is what an application needs to return: a transcript, a translation or a transcript with speaker labels and timing. The models do not share an identical language list or output format.

The release names granite‑speech‑5.0‑2b, granite‑speech‑5.0‑470m‑turboctc and granite‑speech‑4.1‑2b‑plus. Their model cards provide the task and runtime distinctions developers need before choosing an integration.

Granite Speech 5.0 2B lists speech‑to‑text support for English, French, German, Spanish, Portuguese and Japanese. Its translation tasks run between English and those other languages, with additional English‑to‑Italian and English‑to‑Mandarin output. Those extra translation destinations should not be read as Italian or Mandarin transcription support.

IBM also describes input‑language prompting and keyword lists of up to 2,000 terms for transcription and translation. That lets an application supply names or specialised vocabulary as context; it does not establish that every supplied term will be recognised correctly.

The card recommends splitting longer recordings into segments of up to about two minutes before decoding. This is usage guidance, rather than a promise about recognition quality for a particular recording.

The 470‑million‑parameter TurboCTC card lists English, French, German, Spanish, Portuguese, Italian and Japanese transcription. It outputs punctuated and capitalised text. Its stated task is multilingual speech‑to‑text, rather than the translation and speaker‑label modes described for the other variants.

The language distinction is useful for an Italian transcription workflow: Italian appears in TurboCTC’s supported ASR list, while the new 2B model identifies Italian as an additional English‑to‑Italian translation destination. Parameter count alone does not answer which model supports the required task.

Granite Speech 4.1 2B Plus is the option whose card describes speaker‑attributed transcripts and word‑level timestamps. Its supported language list is English, French, German, Spanish and Portuguese. The card explicitly says Plus does not provide punctuation and capitalisation.

Speaker labels are numbered in order of appearance in the recording. That output format is different from a plain formatted transcript and needs to be handled by the consuming application. The card also documents incremental decoding using a previous transcript as context.

The Plus model retains its 4.1 name; IBM describes this as an October update to its weights and configuration, rather than a newly named 5.0 Plus model.

The cards specify different native Transformers requirements: 5.0 2B lists version 4.52.1 or later, TurboCTC lists 5.16.0 or later, and Plus lists 5.8 or later. The new 2B card includes recipes for vLLM, llama.cpp and mlx‑audio; Plus gives a separate vLLM minimum of 0.23.0. A runtime example for one variant should not be assumed to configure another.

All three cards list Apache 2.0. IBM’s 2B cards also warn that generated responses can be inaccurate, biased or unwanted. These task, language and setup details are issuer documentation, not an OpenTools model test or a ranking of transcription quality.

The entrance to IBM’s Thomas J. Watson Research Center in Yorktown Heights, New York, photographed on February 23, 2017. Archive research‑campus context. “IBM Watson Research Center entrance.” Photo: Autopilot, via Wikimedia Commons, CC BY‑SA 4.0 International. Resized and cropped for display; the displayed photo uses the same licence.