Granite Speech 5.0 and 63-language Meet translation: transcription moves to your own server

IBM released three speech models under the Apache 2.0 license that transcribe and translate speech on your own hardware, and Google turned on translation between 63 languages in Meet. For a business owner this means transcription of calls and meetings is no longer a separate paid service. You can run it on your own server, pay nothing per minute, and keep client conversations inside the company.

What came out this week

In a Hugging Face blog post on October 21, the IBM Granite team described three updates. The new granite-speech-5.0-2b transcribes English, French, German, Spanish, Portuguese and Japanese and translates between them and English. You can hand it a list of up to 2,000 terms so it doesn't mangle product names and surnames. By average WER on English benchmarks it ranks second among open models, at 5.06%. The small granite-speech-5.0-470m-turboctc became multilingual, adds punctuation on its own and got about 1.6 times faster, up to 19,870 RTFx. And granite-speech-4.1-2b-plus was tuned for recordings with several speakers: the error rate on long recordings dropped by 24%. All three run in Transformers, vLLM and llama.cpp, and on a Mac through mlx-audio. Russian is not on the language list, and for me that's the main drawback of the release.

On October 22, Google announced in the Workspace Updates blog that speech translation in Meet works between any of 63 languages in a single meeting. The language is detected automatically, and each participant turns translation on for themselves. The limits are noticeable: 90 minutes per meeting on basic Workspace plans and Google AI Pro, 8 hours on AI Expanded Access and Ultra. There is no translation in meeting recordings, phone calls, or meetings with client-side encryption. Google says it doesn't store the audio or train on it. The rollout started on October 22.

On October 14, HubSpot published a roundup of transcription services on its sales blog. It makes one point, and it's a fair one: transcription exists for what comes after it. A summary, action items, who said what, a note in the CRM record. The author says the accuracy claims rest on reviews from G2 and Capterra and advises running one of your own real calls through a service before buying. I agree with that advice.

And one technical note on why a model of your own gets cheaper over time. On October 21, the Dharma-AI team explained on the Hugging Face blog how they rebuilt the tokenizer vocabulary of a Qwen3-4B model that grades essays in Portuguese for them. The vocabulary of 151,669 entries was designed for every language, while the model wrote in one. After the rebuild the output layer shrank from about 152,000 to 65,000 rows, and the measured inference cost fell by 77%. The paper is still under review, so the numbers may change. But when the model runs on your hardware, you can pull levers like this yourself.

Why this matters to a business owner

Take a car rental company. A manager takes dozens of calls a day: extend a rental, report a scratch, ask about insurance. A week later a client says he was promised a discount, and the manager doesn't remember it. If the call was recorded and transcribed, the dispute is settled in a minute: open the text, find the phrase. Paying a service per minute at that volume is expensive, and many owners don't want to hand client conversations to a third party. A model on your own server removes both concerns.

Meetings are similar. Negotiations with a supplier in another country can now run in Meet without an interpreter. But the recording won't contain the translation, so the meeting notes will have to be written separately.

How I have it set up

In late 2023 and early 2024, I built a subtitle system for a consulting company in London on a local Whisper model. It took videos from a Google Drive folder, podcast and YouTube links, extracted the audio, recognized the speech and produced an SRT file with timestamps. There was nothing to pay per minute of recording: the server cost money, but there was no separate charge for volume. Burning subtitles into the video was planned, but the client stopped the project.

At the PDR company in the US, where I've run the server and telephony since 2021, the system records every call, and AI sorts out who was calling (a lead, an existing customer or an insurer) and how the employee handled it. We always know who called, when and why. How the analysis works inside is something I don't disclose. I haven't tested the Granite models myself, but I would test them on exactly those recordings: phone audio, noise, two voices.

What I would do

I'd collect 20 to 30 real recordings: with noise, with accents, with two people talking at once. I'd run any model or service on them before paying.

I'd separate recognition from analysis. First the speech becomes text, then a separate model answers the questions: who called, what was promised, whether a callback is needed. Then the recognition model can be swapped without touching the rest.

I'd decide in advance where the recordings and transcripts live and who can see them. The text of a conversation with a client is their personal data, and it needs the same access rights as their record in the CRM.

I'd run the numbers for a year. At a dozen calls a day, a per-minute service is cheaper than your own server; at hundreds of calls and a steady stream of meetings, your own model pays for itself. I'd name the exact line after counting the volume.

Related reading

How I check model output is described in the article AI in data processing: extraction and verification. The approach: AI in a process where the result can be verified and Python for documents, data and repeatable processing. Cases: SRT subtitles from video on a local Whisper and technical partner of a PDR company in the US. From the digests on the same topic: who called and why: AI call analysis, why I build on free tools and how I choose between my own model and an API. Earlier I wrote that the transcription tool and the model are tested separately. If you have call or meeting recordings piling up, tell me what you need to pull out of them. Discuss your task.