In-depth analysis · About 5 min read
Granite Speech 5 separates accuracy, throughput and speaker attribution
IBM updates three speech models, but language coverage, throughput and speaker-aware errors require different readings of its evidence.
- Published here
- Source published
- Source checked
IBM’s October 21 announcement introduces Granite Speech 5.0 2B and revises Turbo CTC and 4.1 2B Plus. The practical choice concerns the output a workflow needs: a multilingual transcript or translation, fast transcription, or text tied to speakers and timestamps. The results below belong to IBM’s article, not an AISubHub model test.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The new 2B model supports six ASR languages; Turbo CTC lists seven. Translation into Mandarin does not establish Mandarin ASR.
- Turbo’s 19,870 RTFx is benchmark throughput, while Plus reports separate timestamp, word and speaker-related errors.
Start with the output, rather than the version number
The three names describe different jobs. granite-speech-5.0-2b is a new speech-language model for accurate transcription and translation. The updated 470M-parameter granite-speech-5.0-470m-turboctc is an encoder aimed at high throughput and on-device transcription. Neither label makes the other model redundant: processing audio quickly and producing a translated transcript are different requirements.
The updated granite-speech-4.1-2b-plus targets meetings, support calls and other long recordings with several speakers. IBM retains the 4.1 name because its architecture, prompt format and API stay unchanged; it presents the weights as a drop-in replacement. Our analysis: a meeting-search workflow may care more about locating a speaker’s words than a short-form transcription average, so the three products should not be collapsed into one quality ranking.
Language directions and vocabulary controls have boundaries
The 2B model transcribes English, French, German, Spanish, Portuguese and Japanese. Its stated translation directions connect English with French, German, Spanish, Portuguese and Japanese in both directions, plus English into Italian and Mandarin. The latter is a translation-output capability: it does not add Mandarin to the ASR list or establish the reverse direction. Turbo’s seven-language ASR list adds Italian to the six languages above.
IBM reports 390,000 hours of public training audio, with additional synthetic data for Japanese ASR, keyword biasing and translation. It also describes language-ID prompting, requested punctuation and capitalization, and keyword lists of up to 2,000 terms for transcription and translation. Biasing guides recognition toward names or acronyms; it is not a guarantee that a supplied term was spoken. The Plus comparison later uses 20-term lists, not the 2,000-term maximum.
The Conformer change is a throughput claim
IBM says it compiles one Conformer block and reuses its graph across the stack, reducing warmup and raising Turbo CTC from roughly 12,600 to 19,870 RTFx at the same reported accuracy level, about 1.6×. RTFx is an inverse real-time factor: it expresses processing throughput relative to audio duration under the benchmark setup. It is not a measurement of one request’s response delay.
The distinction matters for deployment. A service can have high sustained throughput while a user still waits for buffering, startup, batching or a network round trip. The article’s throughput comparison does not establish those end-to-end delays or a guaranteed speed on a particular device. Our interpretation is to treat the graph-reuse explanation as an optimization mechanism to investigate, not to promise a 1.6× faster interactive conversation.
Read the Plus table as several different errors
For 4.1 2B Plus, IBM’s all-speech average timestamp shift, AAS, changes from 210.86 to 98.54; its short-speech row changes from 29.27 to 26.20. The visible table does not state an AAS unit, so these figures must not be relabeled milliseconds. Long-form WER changes from 13.53 to 10.22, word diarization error rate WDER from 4.40 to 4.10, and speaker-attributed cpWER from 30.50 to 28.48.
The 20-term biased-word WER comparison is 11.79 to 6.48. IBM summarizes the long-form and biased-word improvements as approximately 24% and 45%, and the speaker-related improvements as roughly 7%. These are relative reductions, not percentage-point falls: for example,13.53−10.22 is 3.31 points on that reported WER scale. Different rows answer different questions, so improved timing cannot substitute for checking whether the words or their speaker assignments are correct.
Keep the chart conflict and deployment limits visible
IBM’s static English short-form chart reports 5.06% average WER for the 2B model. Its caption calls the results official as of October 21, whereas the embedded comparison’s footer calls its results unofficial as of October 15 and limits them to public English short-form test sets and open-weight models; the iframe describes nine sets. We preserve those conflicting date and status labels. Reading the chart does not independently verify a current official leaderboard or establish multilingual or long-recording quality.
IBM states Apache 2.0 licensing for all three models and native Transformers support; for the 2B model it also names vLLM, llama.cpp and mlx-audio on Apple Silicon. Our analysis is to select the required output first, then check representative accents, noise, vocabulary, timestamps and speakers in the intended stack. The license does not guarantee hosting costs, production-data permissions, privacy or service availability. We did not run the models, inspect linked model cards or test the browser demo, and the announcement supplies no general production cost or regional-access guarantee.
Why it matters
AISubHub analysis: the useful change is a clearer choice among transcript quality, processing capacity and speaker-aware retrieval. A local evaluation should keep those outcomes separate rather than replace them with one headline score.
Limits and availability
This is an explanation of an IBM-authored release, including its internally inconsistent chart labels. It is not an independent benchmark, a current ranking, a production-price quote or proof of Mandarin ASR.
Sources and evidence
Source published: · Source checked:
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.