Otto365 AI Brief #025
IBM: Granite speech models transcribe calls on your own computer
What happened
IBM's Granite models transcribe your calls and meetings on your own computer. IBM announced Granite Speech 5.0 on the Hugging Face blog on the 21st of October 2026. Three open models transcribe, translate, and label who said what in meetings and calls.1
Three things to know
- Transcribe, translate First, the new 2B model transcribes English, French, German, Spanish, Portuguese and Japanese, and translates to and from English. Give it up to 2,000 keywords, like product names and acronyms, to steer the transcript.1
- Turbo CTC: on-device Second, the Turbo CTC model is built for on-device speed, and it is multilingual with punctuation built in. IBM says it is about 1.6 times faster than before, with a demo in your browser.1
- Who said what, when Third, the 4.1 2B Plus model is tuned for meetings and calls with several speakers. IBM says it labels speakers more accurately, cuts timestamp shift by over half, and makes 24 percent fewer errors on long recordings.1
Why it matters to you
Otto365's take (opinion)
So, my take. Free, open models like these turn calls and meetings into notes without the audio leaving your business. Someone technical must set them up, and check you have consent before you record a call.1
Do this week
Try this: run IBM's browser demo on a short recording.
Sources
- Hugging Face: Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster (opens in a new tab) Primary source, accessed 22 Oct 2026
The quotes we checked (25)
Enterprise Article Published October 21, 2026
Hugging Face · source 1
Results are official, as of 21 October 2026, and averaged over
Hugging Face · source 1
We're excited to share three updates to IBM's Granite Speech collection: a new speech-language model, granite-speech-5.0-2b, and revised versions of two models the community already uses, granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus.
Hugging Face · source 1
Together they cover the full range of enterprise speech work: transcription and translation with high accuracy, extremely fast transcription at the edge, and recordings where who spoke, and when, matter as much as what was said.
Hugging Face · source 1
Granite Speech 5.0 2B: more accurate, more robust, more control
Hugging Face · source 1
It transcribes English, French, German, Spanish, Portuguese, and Japanese, and translates to and from English in those languages, plus English to Italian and Mandarin.
Hugging Face · source 1
Keyword lists up to 2,000 terms: Bias transcription and translation toward product names, people, and acronyms, in every supported language.
Hugging Face · source 1
Language ID prompting: Name the input language in the prompt to steer multilingual transcription.
Hugging Face · source 1
More accurate, more robust English ASR: Higher accuracy on English short-form benchmarks, and better handling of noise, reverberation, and distant-microphone speech.
Hugging Face · source 1
The model ships under Apache 2.0 with support in Transformers, vLLM, llama.cpp, and mlx-audio for Apple Silicon.
Hugging Face · source 1
Granite Speech 5.0 Turbo CTC: now multilingual, more robust, and even faster
Hugging Face · source 1
When we introduced Turbo CTC in August, it was an English-only, 470M-parameter encoder built for speed.
Hugging Face · source 1
Multilingual transcription: English, Japanese, Spanish, German, French, Italian, and Portuguese.
Hugging Face · source 1
Native punctuation and capitalization: Transcripts come out ready to read, with no separate punctuation or truecasing step.
Hugging Face · source 1
pushed Turbo CTC to 19,870 RTFx on the Open ASR Leaderboard, up from roughly 12,600 RTFx at launch: about 1.6x faster, at the same accuracy level.
Hugging Face · source 1
Need raw speed or on-device transcription? Use granite-speech-5.0-470m-turboctc, and try it in the browser with our WebGPU streaming demo.
Hugging Face · source 1
More robust in difficult acoustics: Training that explicitly targets reverberant audio makes Turbo CTC more robust in echoey rooms and far-field recordings.
Hugging Face · source 1
Granite Speech 4.1 2B Plus: better at who said what, and when
Hugging Face · source 1
The updated granite-speech-4.1-2b-plus is tuned for the recordings enterprises actually have: meetings, support calls, and interviews with several speakers running long.
Hugging Face · source 1
Speaker attribution: More accurate labeling of who is speaking. Word diarization error rate (WDER) drops 7%, and speaker-attributed WER (cpWER) drops 7% overall.
Hugging Face · source 1
Timestamps: Average timestamp shift (AAS) is cut by more than half, yielding tighter timing in captions, search, and review.
Hugging Face · source 1
Long-speech performance: WER on long-form test sets drops 24%.
Hugging Face · source 1
it's a drop-in replacement for the current model with improved weights.
Hugging Face · source 1
These models are all released under an Apache 2.0 license with native support in Transformers.
Hugging Face · source 1
We recommend pairing them with Granite Guardian for production deployments that need additional risk detection.
Hugging Face · source 1
Transcript
Full transcript
IBM's Granite models transcribe your calls and meetings on your own computer.
Play from 0:04 IBM announced Granite Speech 5.0 on the Hugging Face blog on the 21st of October 2026. Three open models transcribe, translate, and label who said what in meetings and calls.
Play from 0:18 First, the new 2B model transcribes English, French, German, Spanish, Portuguese and Japanese, and translates to and from English. Give it up to 2,000 keywords, like product names and acronyms, to steer the transcript.
Play from 0:34 Second, the Turbo CTC model is built for on-device speed, and it is multilingual with punctuation built in. IBM says it is about 1.6 times faster than before, with a demo in your browser.
Play from 0:48 Third, the 4.1 2B Plus model is tuned for meetings and calls with several speakers. IBM says it labels speakers more accurately, cuts timestamp shift by over half, and makes 24 percent fewer errors on long recordings.
Play from 1:04 So, my take. Free, open models like these turn calls and meetings into notes without the audio leaving your business. Someone technical must set them up, and check you have consent before you record a call.
Play from 1:18 Try this: run IBM's browser demo on a short recording.
I'm Otto365.
Follow the Brief.
On screen
- Play from 0:00
- Otto365 AI Brief #025, IBM Granite, 21 Oct 2026
- Granite transcribes calls on your own computer
- Your call
- Your own computer
- Transcript
- Who said what
- Play from 0:04
- Granite Speech 5.0
- Announced 21 Oct 2026
- Three open speech models
- 5.0 2B: transcribe, translate
- Turbo CTC: on-device speed
- 4.1 2B Plus: who said what
- Source: Hugging Face, 21 Oct 2026
- Play from 0:18
- Transcribe, translate
- Your recording
- Set the language
- Keyword list
- Transcript
- Product names
- People
- Acronyms
- Up to 2,000 terms
- Play from 0:34
- Turbo CTC: on-device
- Before: English only, Punctuation step
- Updated: Multilingual, Ready to read
- 1.6x Faster, same accuracy
- Play from 0:48
- Who said what, when
- Support call transcript
- Speaker labels
- Timestamps
- Long recordings
- Keyword biasing
- Speaker labels: more accurate
- Timing: shift cut by over half
- Long audio: 24% fewer errors
- Play from 1:04
- Audio stays in-house
- Call recording
- Your own computer
- Who said what
- Search and review
- Apache 2.0 licence
- Transformers, vLLM, llama.cpp
- Pair with Granite Guardian
- Play from 1:18
- Run the browser demo
- Voiced by AI. Checked by humans. Apps 365 sells AI and Microsoft 365 services. Not endorsed by companies named.
Voiced by AI. Checked by humans. Apps 365 sells AI and Microsoft 365 services. Not endorsed by companies named. Report an error.