IBM Granite: transcribe speech, plan your stock

Turning a voice message into text and preparing a stock order: these are two concrete tasks addressed by IBM’s latest announcements. For a small team, their appeal is the chance to consider a focused trial. Before changing how you work, choose a task whose result you already know how to check.

From voice notes to usable text

On October 21, IBM presented a new Granite Speech model and two updates. The announcement separates transcription, translation and identifying who is speaking: these functions are not interchangeable. In a meeting, knowing who promised what can matter as much as getting the words right. IBM announcement.

The TurboCTC version now supports French among seven languages and produces transcripts with punctuation and capitalization. Its model card lists an Apache 2.0 license and targets uses including execution on local devices. This is still a model to integrate, rather than a finished note-taking application you can simply open. Model card.

Think of it as a secretary’s first draft: you hand over an audio note, then correct the text before sending it. The comparison stops there: do not assume the tool understands your business or checks a promise. For an initial trial, use a recording you have permission to process, including product names you know. Pay particular attention to amounts, dates and negatives: these are useful places to look for an error that would change the meaning.

Forecast demand without relying on one model

On October 22, IBM Research presented Granite Timeseries Ensemble. Its default recipe combines four previously trained models to produce a central forecast and ranges, without new training on your data. One possible experiment would be weekly sales. IBM Research introduction.

A useful comparison is several colleagues each preparing an estimate before pooling their answers. Four opinions do not guarantee a good answer. Here, the models calculate from historical records; they are not people who can spontaneously notice that a supplier is closing or a promotion is coming.

The model card explains that the recipe downloads the models separately and can continue if one member fails, changing the combination. Check what actually ran before comparing the results. Recipe card.

A useful trial starts with a specific question

We have run neither these speech models nor this forecasting recipe. The announcements and model cards are available to read; the reported quality improvements remain supplier results. They do not establish a benefit for your business.

For stock planning, choose a past period, hide the sales that follow and request a forecast. Then compare it with actual sales and your usual method over the same period. For speech, count the corrections needed before someone else can use the text.

Decide beforehand what would justify continuing: fewer corrections, a better prepared order or work that is easier to review. If the trial does not answer that question, you will still have learned where the tool helps and where it needs too much supervision.