IBM’s 470-million-parameter Turbo CTC model now transcribes seven languages, expanding beyond English. Alongside it, a new Granite Speech 5.0 2B speech-language model handles transcription, translation and vocabulary prompting. Builders working on meetings and calls get a separate option: an updated Granite Speech 4.1 2B Plus model focused on speaker attribution and timestamps.
The October 21 announcement describes three deployment choices. All three carry Apache-2.0 licenses and native Transformers support. The new 2B model also supports vLLM, llama.cpp and mlx-audio for Apple Silicon.
A transcription service processing large audio queues has different requirements from a translation application or a meeting assistant that must preserve who said what. Those requirements are more useful guides than the version numbers.
| Model | Strongest reason to evaluate it | Important boundary |
|---|---|---|
granite-speech-5.0-2b | Multilingual transcription, translation and large keyword lists | Its six-language transcription coverage is not identical to its translation coverage |
granite-speech-5.0-470m-turboctc | High-throughput or on-device transcription across seven languages | Its encoder-only design gives up speech translation and keyword biasing |
granite-speech-4.1-2b-plus | Long recordings, speaker attribution and tighter timestamps | It remains a separate 4.1 model, not a speaker-aware mode of the new 5.0 2B model |
The 2B Model Handles Translation and Vocabulary Control
The new Granite Speech 5.0 2B model succeeds Granite Speech 4.1 2B at the same stated parameter count. IBM says it improves English short-form recognition, robustness to noise and reverberation, and translation of longer, conversational recordings.
Its language coverage needs careful reading:
- Transcription: English, French, German, Spanish, Portuguese and Japanese.
- Bidirectional translation: Between English and the other five transcription languages.
- Additional translation directions: English to Italian and English to Mandarin.
Italian and Mandarin therefore should not be treated as additional supported transcription languages for this model. The announcement also does not establish unrestricted translation between every pair of supported languages.
Developers can supply lists of up to 2,000 terms to steer transcription and translation toward relevant vocabulary, including product names, people’s names or acronyms. IBM says this keyword biasing works across the supported languages.
That makes the model worth evaluating for domain-specific recordings where a small set of unusual words carries much of the value. An otherwise readable transcript can still be unsuitable for search or recordkeeping if it repeatedly misrecognizes the product or person under discussion. Keyword biasing can steer the output, but it does not guarantee that supplied terms will be recognized correctly.
The model also accepts an input-language identifier in the prompt and can produce punctuation and capitalization on request. These controls distinguish its speech-language-model approach from a transcription-only encoder. Its LLM component supports prompted audio-to-text tasks, though that should not be confused with a claim that it is a general-purpose audio assistant.
IBM reports an average English word error rate of 5.06%, placing it second among the open-weight ASR models in the announcement’s comparison. That result averages the Open ASR Leaderboard’s public English short-form test sets as of October 21. It does not establish translation quality, multilingual accuracy or reliability on hour-long meetings; those require separate evaluations.
Turbo CTC Adds Seven-Language Transcription, Not Translation
The revised Turbo CTC model keeps its 470M-parameter footprint while expanding beyond English. It now transcribes English, Japanese, Spanish, German, French, Italian and Portuguese.
It also adds native punctuation and capitalization, alongside training intended to improve recognition in reverberant rooms and distant-microphone recordings. For applications that need readable transcripts, native formatting can remove the need for a separate punctuation and capitalization stage.
IBM’s original Turbo CTC announcement describes an encoder-only model, unlike earlier Granite Speech systems that combined an acoustic encoder with a language model. That design targets transcription speed and a smaller parameter footprint, while giving up capabilities such as speech translation and keyword biasing.
The updated model reaches 19,870 RTFx, according to IBM, compared with roughly 12,600 RTFx at launch. IBM describes that as about 1.6 times faster at the same accuracy level, attributing the improvement to compiling one Conformer block and reusing its graph across the stack. The company also reports reduced warmup time.
RTFx expresses throughput as audio duration divided by processing time. Higher is better, but it does not measure the delay before a live application displays its first words.
The launch result was measured with batched inference on an NVIDIA H200 GPU. The new figure comes from IBM’s reported Open ASR Leaderboard results, not a laptop benchmark. Neither figure establishes what a browser, CPU or Apple Silicon deployment will achieve.
Aggregate throughput is highly relevant to a batch-transcription service. Live captions may depend more on first-result latency, update frequency and performance with a single audio stream. A model can perform well on one measure without providing the best experience on the other.
The 4.1 Plus Update Targets Errors That Break Meeting Tools
The updated Granite Speech 4.1 2B Plus keeps its existing architecture, prompt format and API. IBM describes it as a drop-in replacement with improved weights, aimed at long, multi-speaker recordings such as meetings, interviews and support calls.
Its reported evaluation results show improvements across several distinct failure modes:
| Metric | Previous model | Updated model |
|---|---|---|
| Long-form word error rate | 13.53 | 10.22 |
| Biased-word error rate with 20-term lists |
Sources
- October 21 announcementhuggingface.co
- Granite Speech 5.0 2B modelhuggingface.co
- original Turbo CTC announcementhuggingface.co
