{
  "id": 17545068,
  "title": "Nemotron 3.5 Transcribe: One-GPU Promptable Offline STT for Multilingual Call Pipelines",
  "url": "https://urgent.news/2026/11/01/nemotron-3-5-transcribe-one-gpu-promptable-offline-stt-for",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-01T06:14:55.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/madnessbrains_i/nemotron-35-transcribe-one-gpu-promptable-offline-stt-for-multilingual-call-pipelines-1ejd"
  },
  "original_language": "en",
  "account": "NVIDIA introduced Nemotron 3.5 Transcribe on October 28, 2026 as a single-weight, offline speech-to-text (STT) solution supporting 40 languages, translation to/from English in 31 of them, speaker identification for up to eight speakers in 13 languages, language detection, word-level timestamps, and audio question answering—all controllable via natural-language prompts. Offline-only, it processes complete recordings up to 65,536 tokens (one hour). The checkpoint, named 30B-A4B, contains approximately 34 billion parameters, with about 4.3 billion active per token. The model architecture combines Nemotron 3.5 Lightning hybrid Mamba-2/Transformer mixture-of-experts with a parallel ASR encoder and frozen Nemotron 3 diarization. NVIDIA claims the NVFP4 version can handle up to 2,238 seconds of audio per second on an RTX PRO 6000 Blackwell GPU, enabling 100-hour short clips in under three minutes. Ranked first on the Hugging Face Open ASR English leaderboard with an average Word Error Rate (WER) of 5.33%, Nemotron 3.5 Transcribe demonstrates strong performance in independent benchmarks as well. Fine-tuning on domain-specific data (e.g., medical, earnings calls) further reduces errors. A Windows tool can buffer 5-15 seconds of audio captured via WASAPI and feed the transcription results directly into downstream LLMs or an interactive UI. However, the model is not suitable for true low-latency streaming; for that, other options such as Nemotron 3.5 ASR streaming 0.6B or local Whisper alternatives are recommended. The model is built on an NVIDIA GPU, with quantized GGUF and community ports available, but optimal performance relies on Blackwell or equivalent hardware. Speaker diarization performs well on the supported 13 languages, while other languages receive transcription and translation but less reliable labels. Audio question answering exhibits strongest performance in English. Implementers can serve the model via vLLM, Hugging Face Transformers, or the NeMo Speech stack on Windows. Community ports to llama.cpp-style runtimes (e.g., GGUF) have emerged quickly. Pairing Nemotron 3.5 Transcribe with a fast capture method (WASAPI exclusive or shared low-period, or loopback) and a lightweight voice activity detector (VAD) ensures optimal performance. The model's promptable interface enables flexible output styles per call type without retraining, making it an efficient alternative to stitching together multiple separate models. Granary v2, the accompanying open dataset, includes 463,000+ hours of audio in 48 languages and three transcript styles, offering further customization. Overall, Nemotron 3.5 Transcribe represents the most robust open-source offline STT solution for multilingual, speaker-aware applications on NVIDIA GPUs, particularly well-suited for Windows-based call and interview tools.",
  "summary": "NVIDIA released Nemotron 3.5 Transcribe on 28 October 2026. The open-weight speech-augmented language model (SALM) handles transcription in 40 languages, translation to or from English in 31 of them, speaker attribution for up to eight speakers in 13 languages, language identification, word-level timestamps, and audio question answering—all from a single checkpoint you steer with a…",
  "key_points": [],
  "editors_take": "Consolidating transcription, translation, and diarization into a single promptable model streamlines multilingual call processing workflows by removing the need to stitch together multiple separate speech models.",
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}