NVIDIA Releases Nemotron 3.5 Transcribe, Using Natural Language Instructions to Control Multilingual Transcription

On October 28, NVIDIA released Nemotron 3.5 Transcribe, which lets developers use text instructions to choose whether a recording should produce a verbatim transcript, normalized written text, or a transcript with speaker labels. The same model can also handle translation and audio question answering, consolidating work that was previously spread across multiple processing stages into a single model. It is designed for offline audio processing, not streaming speech recognition. Official release notes

The architecture connects a Transformer speech encoder and a frozen speaker diarization model to the Mamba–Transformer Mixture-of-Experts backbone of Nemotron 3.5 Lightning. Including the audio encoder, it has about 34 billion parameters, with about 4.3 billion active per operation. Speaker activity signals are injected into the acoustic representations to help generate text with speaker labels; a CTC alignment module then produces word-level timestamps. This allows output format control and multi-speaker recognition to be handled within a shared architecture. Technical overview

Capabilities vary by task: the model transcribes 40 languages, supports translation between English and 31 languages, and supports speaker attribution in 13 languages, with up to eight speakers labeled per recording. Audio question answering was trained in English. For Chinese-language applications, “supports 40 languages” should not be taken to mean that every feature has the same language coverage; check the task matrix for each capability. NVIDIA announcement

For deployment, the model card provides an example for running on a single 80GB H100 and lists a 65,536-token context window. NVIDIA claims that with NVFP4 and DFlash2 speculative decoding, a single RTX PRO 6000 Blackwell can process up to 2,238 seconds of audio per second. However, the model card states that these NIM serving optimizations are not included in the open-source release, so the throughput figure cannot be applied directly to self-hosted deployments. Model card

Engineering teams should next evaluate their own recordings for proper nouns, overlapping speech, and omissions in long recordings, while measuring batch throughput separately from per-request latency. Open weights provide room for customization and self-hosting, but actual costs still depend on the full weights, numerical precision, and serving implementation; they cannot be estimated from active parameter count alone.