scribe — local transcription & speaker diarization on NVIDIA GPUs (Linux)

Acknowledgements & gratitude

This project is a fork of scribe by Trail of Bits. The original repo, its design, and the great majority of its code are entirely their work — without it this NVIDIA-centric variant would not exist.

Please go star and credit the original repository: https://github.com/trailofbits/scribe. This fork only adapts that work to run natively on NVIDIA GPUs (including the NVIDIA DGX Spark), by swapping the Apple/MLX transcription backend for NVIDIA NeMo + Parakeet and the macOS-native diarization for the senko GPU stack. All credit for the overall architecture, the parallel-loading pipeline, the merging logic, the CLI, the tests, and the documentation belongs to the original authors. Thank you, Trail of Bits.

---

Quick install

uv tool install https://github.com/eleqtrizit/nvscribe.git

Then run scribe from anywhere. See Requirements below for prerequisites (NVIDIA GPU, Linux, ffmpeg).

aarch64 / DGX Spark: uv tool install from a git URL does not apply [tool.uv] override-dependencies from pyproject.toml, so the triton platform override is ignored. Use an overrides file instead:

echo 'triton; platform_machine == "x86_64"' > overrides.txt
uv tool install --overrides overrides.txt https://github.com/eleqtrizit/nvscribe.git

---

Local transcription and speaker diarization running entirely on NVIDIA GPUs on Linux, built on NVIDIA's own software stack:

• Transcription — NVIDIA NeMo with the parakeet-tdt-0.6b-v3 FastConformer-TDT ASR model.
• Diarization — senko, which runs Pyannote segmentation and CAM++ speaker embedding models on the CUDA GPU via PyTorch.

⚠️ Tested hardware: This fork has been tested on the NVIDIA DGX Spark (Grace Blackwell GB10, aarch64) and the NVIDIA GeForce RTX 5090 (Blackwell GB202, x86_64). It targets NVIDIA hardware and an NVIDIA CUDA driver. Other NVIDIA GPUs (Ampere, Hopper, Blackwell datacenter cards, etc.) should work but are untested by the maintainer. Non-NVIDIA platforms (Apple Silicon, CPU-only, AMD ROCm, etc.) are not supported, and macOS is not supported — this fork requires Linux + an NVIDIA CUDA driver. Use the original scribe for macOS / Apple/MLX.

Takes any audio or video file ffmpeg can read (mp4, mp3, m4a, webm, …), transcribes the speech with Parakeet TDT v3, optionally identifies who spoke when using GPU diarization, and produces a speaker-attributed transcript. Input is automatically converted to 16 kHz mono WAV for processing. Transcription and diarization both run on the CUDA GPU, and models download automatically on first run from Hugging Face (no account needed).

Requirements

• An NVIDIA GPU with a CUDA-capable driver. Developed and tested on the NVIDIA DGX Spark (Grace Blackwell GB10) and the NVIDIA GeForce RTX 5090 (Blackwell GB202). CUDA compute capability ≥ 8.0 (Ampere / Hopper / Blackwell) is expected.
• Linux — this fork does not run on macOS or Windows. The DGX Spark ships Ubuntu; other Linux distros may work but are untested.
• Python 3.11+
• uv
• ffmpeg — all input is normalized to 16 kHz mono WAV: apt install ffmpeg

Install

uv sync

This installs nemo_toolkit[asr] (which pulls in a CUDA-enabled PyTorch from download.pytorch.org/whl/cu129) plus senko. On first run the Parakeet and diarization models are downloaded to the local Hugging Face / NeMo cache.

Note for DGX Spark (aarch64): see aarch64 / Grace Blackwell notes below for why the senko[nvidia-windows] extra and a triton platform override are used in pyproject.toml.

Usage

scribe meeting.wav
scribe recording.mp4
scribe podcast.mp3

Writes meeting.txt / recording.txt / podcast.txt:

SPEAKER_00: Thank you for joining. Let's start with the agenda.

SPEAKER_01: Sure, I wanted to discuss the roadmap first.

Transcript only (no diarization)

scribe meeting.wav --no-diarize

Options

scribe meeting.wav -o transcript.txt      # custom output path
scribe meeting.wav -o -                   # write to stdout
scribe meeting.wav --format json          # JSON with timestamps
scribe meeting.wav --format json -o -     # JSON to stdout
scribe meeting.wav --no-diarize           # transcript without speakers

JSON output

scribe meeting.wav --format json

Writes meeting.json with timestamps and speaker attribution:

{
  "speakers": ["SPEAKER_00", "SPEAKER_01"],
  "segments": [
    {
      "speaker": "SPEAKER_00",
      "text": "Thank you for joining.",
      "start": 0.531,
      "end": 2.874
    }
  ]
}

Notes on NVIDIA hardware / memory

The Parakeet encoder runs with local attention (rel_pos_local_attn, context 256+256) so long-form audio (up to ~3 hours) transcribes with a small, predictable memory footprint — important on shared GPUs like the DGX Spark. If your GPU is busy with other workloads, transcription and diarization still run in parallel on the same device.

aarch64 / Grace Blackwell (DGX Spark) notes

The DGX Spark's GB10 is aarch64. Two packages in senko's full GPU stack lack aarch64 wheels, so pyproject.toml selects workarounds:

• senko[nvidia-windows] instead of senko[nvidia] — this skips kaldifeat and RAPIDS (cudf / cuml), which have no aarch64 wheels. The neural diarization models (Pyannote segmentation, CAM++) still run on the CUDA GPU via PyTorch; only fbank extraction and clustering fall back to CPU (senko's documented behavior when kaldifeat is absent).
• [tool.uv] override-dependencies scopes triton to x86_64, since torch's cu129 wheel requires triton on all Linux but triton has no aarch64 wheel. triton is only needed for torch.compile, not inference.

The CUDA-enabled torch comes from download.pytorch.org/whl/cu129 (torch==2.8.0+cu129), which ships Blackwell-capable aarch64 wheels. If a future torch release drops or adjusts these, re-run uv sync to re-resolve.

On the DGX Spark, torch may print a one-time UserWarning that the GB10 is CUDA capability 12.1 while this build supports up to 12.0 — CUDA still initializes and inference runs correctly; a future torch+cu130 wheel will silence it. A similar one-time onnxruntime warning about /sys/class/drm/card0/device/vendor may appear; it comes from silero-vad's onnxruntime import-time GPU probe (which has no aarch64 CUDA provider and falls back to CPU) and is harmless.

Development

uv sync --dev
uv run pytest -q
uv run ruff check

License

MIT