# Speech: what I've logged

> Speech comes up in 2 posts of my AI log, from October 2026 to October 2026, and each item below links to the primary source. The sources come from NVIDIA and Block, filed under frontier, open weights and science.

## Items

1. **NVIDIA Nemotron 3.5 Transcribe and Granary v2** (NVIDIA, 28 Oct 2026): Open-weight promptable speech model covering transcription, speaker labels, translation and audio Q&A in 40 languages, plus a 463,062 hour multilingual dataset release. [Primary source](https://huggingface.co/blog/nvidia/nemotron-transcribe-granary) · [Post](https://heyaliasif.com/ai/2026-10-28-gemini-4-carbon-nemotron-transcribe-olmocr-ai-for-science#item-2)
2. **Dessa: a conversation model trained only on speech** (Block, 23 Oct 2026): A full-duplex speech-to-speech model trained from scratch on human conversation with no text, said to run faster than real time and locally on an M4 Max MacBook Pro or better. [Primary source](https://block.xyz/dessa) · [Post](https://heyaliasif.com/ai/2026-10-24-dessa-hy-nomas-tensorrt-tuner-microduck#item-1)

## Posts

- 2026-10-28 [Gemini 4 Carbon, Nemotron 3.5 Transcribe, olmOCR 2 archives](https://heyaliasif.com/ai/2026-10-28-gemini-4-carbon-nemotron-transcribe-olmocr-ai-for-science)
- 2026-10-24 [Dessa, Hy-Nomas, TensorRT 11.1 tuner, Microduck](https://heyaliasif.com/ai/2026-10-24-dessa-hy-nomas-tensorrt-tuner-microduck)

Canonical: https://heyaliasif.com/ai/topics/speech