Cách sử dụng NVIDIA Nemotron 3.5 Transcribe miễn phí

Description and use

Nemotron 3.5 Transcribe is a free, high‑accuracy ASR model from NVIDIA that can be run locally on your GPU or via a free Space demo. It supports 40 languages, up to one hour of audio, speaker diarization, word‑level timestamps, and translation.

Problem

Need a free, high‑accuracy speech‑to‑text solution that can be run locally or via free demos.

Outcome

Users can transcribe up to one hour of audio per request, support 40 languages, speaker diarization up to eight speakers, word‑level timestamps, and translation to 31 languages, all for free when self‑hosted or via the demo.

Workflow

• 1. Choose a free route: NVIDIA Space demo, community Spaces, or self‑hosted weights.
• 2. For Space demo: open https://huggingface.co/spaces/nvidia/nemotron-transcribe, select a tab (Karaoke, Podcast, Subtitles), upload or link a sample, run and view output.
• 3. For self‑hosted: install transformers via pip, load model with AutoModelForMultimodalLM and AutoProcessor, prepare audio file, call processor.apply_transcription_request, generate, and decode output.
• 4. For batch or production: deploy with vLLM 0.32.0 and NeMo Speech plugin to expose an OpenAI‑compatible endpoint.
• 5. Monitor GPU usage and choose appropriate precision (BF16, FP8, NVFP4) based on hardware (H100, A100, RTX 5090).

Evidence

Nemotron 3.5 Transcribe (full name NVIDIA‑Nemotron‑3.5‑Transcribe‑30B‑A4B) was released on 28 Oct 2026 under OpenMDW‑1.1, allowing commercial use.

Source

It achieves a mean WER of 5.33% on the Open ASR leaderboard, ranking 9th out of 61 models.

Source

The Space demo is free, no login required, but only previews 60 seconds and shares GPU resources.

Source