IBM Granite Speech 5.0: Choosing Accuracy, Speed or Speaker Attribution

IBM's October 21, 2026 Granite Speech announcement introduces three checkpoints: a new 2B speech-language model, a multilingual update to 470M TurboCTC, and refreshed 4.1 2B Plus weights. The choice depends on what the application needs to preserve beyond the transcript.

A speech application can get the words right and still fail its user. A meeting tool might assign a commitment to the wrong person. A captioning system might place accurate text several seconds late. An on-device transcription app might simply take too long.

The new 2B model adds control over transcription

The Granite Speech 5.0 2B model card dates its release to October 20. It supports transcription in English, French, German, Spanish, Portuguese and Japanese, with speech translation between English and those languages. English-to-Italian and English-to-Mandarin translation are additional paths.

Two controls stand out: specifying the input language and supplying a keyword list of up to 2,000 terms. IBM also reports improvements in difficult acoustics and longer conversational translation.

For a product handling specialist vocabulary, the keyword list is worth evaluating carefully. A transcription that misses one product identifier can be less useful than another with several harmless grammatical errors. A practical test should count critical-name mistakes separately from overall word error rate.

The card documents Transformers, vLLM, llama.cpp and mlx-audio usage. It recommends audio segments of roughly two minutes or less, splitting longer recordings before decoding. For llama.cpp, IBM recommends Q8_0 and warns that lower-bit quantization loses accuracy on audio beyond about a minute.

TurboCTC extends the smaller, faster option

The October 21 TurboCTC update expands the previously English-only 470M model to seven languages, including Italian and Japanese. It now produces punctuation and capitalization directly.

Its CTC architecture decodes non-autoregressively rather than generating a transcript token by token. IBM positions it for high-throughput and edge transcription. The card reports average English word error rate improving from 5.78% to 5.51% on its short-form benchmark comparison, despite the added languages.

IBM's announcement separately reports throughput of 19,870 RTFx, about 1.6 times its launch result. That measures audio processed relative to elapsed processing time under the benchmark setup. It should not be read as an interactive latency promise for a laptop or phone.

The documented Transformers requirement is version 5.16.0 or later. That dependency matters when an application already pins an older runtime.

Plus preserves speaker and timing information

The updated 4.1 2B Plus card lists October 19 as its update date. It retains the older model name and adds value through speaker-attributed transcripts and word-level timestamps. Unlike the base speech model, Plus does not supply punctuation and capitalization.

The timing improvement is substantial in one published test: average word-timestamp shift falls from 814.3 milliseconds to 328.3 milliseconds on long samples assembled from shorter recordings.

There is an important trade-off. On the card's updated English short-form benchmark sets, average WER moves from 6.01% to 6.10%, slightly worse. These are IBM's reported evaluations, not independent tests conducted for this article. Better timing should not be presented as better performance on every speech task.

For a searchable recording archive, tighter timestamps could make jumping to a passage more reliable. A meeting assistant also needs a separate check that the right speaker owns each statement.

Evaluate the output your product actually uses

All three models are available under Apache 2.0. These releases provide downloadable weights rather than a newly priced hosted transcription service. The announcement does not quote a per-minute API price; deployment costs still depend on compute, concurrency and the surrounding pipeline.

I would compare candidates on the same recordings and separate four measurements: transcription errors, important-name errors, speaker mistakes and timing accuracy. Include the acoustics and languages customers actually use, rather than relying only on clean benchmark clips.

Measure the complete workflow too. Audio preparation, segmentation and any formatting step can change the experience more than a headline inference number suggests. Pin the tested model revision so that a later weight update does not silently change the comparison.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn model releases and engineering trade-offs into clear stories that developers can use.

Book a 30-minute call.