IBM expands Granite Speech TurboCTC to seven languages
IBM updated Granite Speech 5.0 TurboCTC on October 21, adding six languages to its previously English only speech recognizer. The 470 million parameter model now produces punctuated and capitalized transcripts in English, French, German, Spanish, Portuguese, Italian and Japanese. Its model card distinguishes this release from the original version published on August 25.
The update is part of a broader Granite Speech announcement covering three models with different jobs. TurboCTC targets fast transcription, the new 2B model adds translation and vocabulary steering, and the revised Plus model focuses on speakers and timing. Choosing among them means deciding what the transcript needs to contain, rather than assuming the largest model is the right default.
What changed in the small transcription model
TurboCTC retains its compact architecture but expands its output vocabulary for multilingual speech. It uses greedy decoding rather than generating a transcript token by token through a language model. IBM documents native support in Transformers 5.16.0 or later and an Apple Silicon path through mlx audio 0.5.1 or later.
IBM also reports better handling of reverberation and distant microphones. Those are useful targets for recordings made outside a studio, but the announcement does not establish accuracy on every accent, microphone or room. A team considering the update should compare actual recordings with manually checked transcripts before replacing an existing system.
There is an important benchmark qualification. The TurboCTC card labels the new multilingual result unofficial and says the leaderboard entry still represents the older English only model. Readers should not mistake a chart containing the new checkpoint for an independently updated leaderboard ranking.
When the 2B model is the better fit
The separate Granite Speech 5.0 2B card dates that model to October 20, not October 21. It supports transcription in six languages and translation between those languages and English, with additional English to Italian and Mandarin translation. Italian transcription belongs to TurboCTC, not this model’s documented transcription list.
The 2B model accepts keyword lists of up to 2,000 terms to help recognize names, acronyms and specialist vocabulary. IBM recommends segments of roughly two minutes or less for best results. Longer recordings should be divided before decoding. Its local deployment guidance includes Transformers, vLLM, llama.cpp and mlx audio.
Quantization needs care. IBM recommends Q8_0 for its llama.cpp recipe and warns that lower precision versions lose accuracy on audio longer than about a minute. That is a practical reason to evaluate a chosen download, rather than transferring a result from the original weights to every smaller variant. The card also warns that outputs can be inaccurate or biased.
Speaker labels are a separate capability
The Granite Speech 4.1 2B Plus update is dated October 19 and provides context rather than a newly released October 21 model. It can add speaker labels and word timing to transcripts. Unlike the base model, it does not supply punctuation and capitalization. Its documented languages are English, French, German, Spanish and Portuguese.
The distinction matters for meeting archives. Speaker labels and timestamps can help locate a statement, but they are not interchangeable with a polished transcript. IBM reports improvements in timing and keyword recognition while its revised short English test table shows average word error rising from 6.01 to 6.10 percent. The update is not a universal accuracy improvement.
The Plus card also explains that benchmark test sets changed. Older published averages should not be compared directly with the current leaderboard. IBM evaluated segments up to nine minutes for transcription and speaker attribution, and up to three and a half minutes for timestamps. Those are reported evaluation ranges, not guarantees for arbitrary recordings.
Availability and a sensible evaluation path
IBM releases all three models under Apache 2.0. The useful comparison is workload specific. Choose plain multilingual transcription, translation with specialist vocabulary, or a record of who spoke and when, then test that requirement directly. ByteForward has not run these models. Reported gains remain IBM’s results, not measurements from our own hardware.
Readers comparing speech releases can also review ByteForward’s Voxtral 3 coverage. Deployment requirements and migration changes deserve attention alongside model quality.
Illustrative studio microphones. Photo by Clement Lepetit on Pexels under the Pexels License.
