On 28 October Nvidia released Nemotron 3.5 Transcribe, an open-weight speech-to-text model that supports 40 languages. It processes recordings of up to an hour, tells speakers apart, can translate and gives word-level timestamps. It works on recorded material rather than live audio. It comes in several versions, including GGUF for running locally, under the OpenMDW-1.1 licence.

What it means: Transcription is one of the most practical business uses of AI: meeting minutes, interviews, customer service calls and medical notes. Until now most good options ran as cloud services, which meant the audio left the company for a third-party provider. An open model that runs on your own hardware changes that equation, especially for law firms, clinics, accounting offices and public bodies that handle sensitive data and must comply with the GDPR. The key question for Greece is quality in Greek, which usually lags behind English, and how the model handles terminology, names and regional accents. Before any production use, it is worth testing on the company's real recordings and comparing with the services already in use. The lack of live audio support also means it does not cover real-time call centres. Finally, running a model of this size needs a graphics card, so the infrastructure cost has to be compared with a provider's subscription fee.

Read the original on NVIDIA

Related service: Custom software

Related stories