feat: VAD-only GGUF files (head, front end, subsampler) and a slicer

[localai-org-maint-bot]

What and why

Ultra and Redux have a built-in VAD head, but using it meant loading the whole model, 213 MB to 1.4 GB. The head reads only the log-mel front end, the subsampler and itself. This PR adds a script that cuts those parts out into a 6 to 10 MB GGUF, and loaders that run VAD from it.

How

• scripts/slice_vad_gguf.py copies 20 tensors byte for byte from an Ultra or Redux GGUF, plus the encoder, preprocessor and parakeet.vad.* keys. It does not requantize. It records the parent (name, file, size, architecture, SHA-256) as provenance metadata.
• The slice sets parakeet.arch=vad. A normal file has a vocabulary and a decoder, and the slice has neither, so the loader needs the marker.
• ModelLoader::load accepts the marker without a vocabulary. New Model::load_vad_only serves it. parakeet-cli vad (via gguf_is_vad_only) and the parakeet_capi_vad_* functions use it through a separate context slot.
• Every other call refuses the slice with a clear message: Model::load, DiarizationModel::load, and the other C-API entry points ("context holds a VAD-only model"). Files that are not slices load as before.

Results

Compared with the parent model on three clips and four slices (Ultra Q8_0, Ultra F16, Redux packed, Redux dequantized F16):

• Output is byte-identical to the parent.
• Load time is a few ms, against 0.1 to 0.7 s.
• Peak memory is about 245 MiB, against 0.6 to 1.6 GiB, for a 33 s clip.
• Speed is unchanged. VAD at 1 thread runs at about 92x real time, against 736x for Silero. The machine was busy, so the 8-thread numbers are only indicative.

I also ran parakeet-cli vad on the two published files below and compared with the full parent output on two_speakers.wav: identical.

Tests

• New tests/test_vad_only.cpp (label model; needs PARAKEET_TEST_VAD_ONLY_GGUF and PARAKEET_TEST_GGUF_ULTRA): the slice gives the parent's VAD output, and the slice is refused by the other calls.
• ctest -LE model: 35 of 35 pass (5 skipped for missing fixtures). With the model GGUFs set, everything passes except test_relpos_attention_local_chunked and test_capi_timestamps, which fail without this change too.
• CED and voice-detect were off in my build because their submodules were not checked out.

Limits

• CPU only. Streaming and GPU were not tested.
• The C ABI is unchanged.
• No accuracy rerun, since the output is identical to the parent.
• A slice cannot transcribe, diarize or stream.
• The bundle work (feat: bundle GGUF format (ASR, VAD, diarization, sound events and voice in one file) #85, if it lands) touches the same loader and C-API code. The two features do not depend on each other: a slice is its own file kind, and a bundle does not need to hold one.

Published files

Both are on https://huggingface.co/mudler/parakeet-cpp-gguf, CC-BY-4.0. They are cut out of Moondream's models, not trained here: credit Moondream and NVIDIA. docs/vad.md, the README and models/MANIFEST.md link them with sizes and checksums.

File	Size (bytes)	SHA-256
redux-vad.gguf	9,939,488	588e1d6e2ee5b6cdfd9ec5ea98dc0993d5bea498d9cc4ec8d6077041eef8a34f
ultra-vad-q8_0.gguf	6,007,328	8b891a4435e97438104ca07c72530d0c5fe62b986baee48b2dd4e1500c1d4758

parakeet-cli vad --model redux-vad.gguf --input audio.wav

Assisted-by: Claude:claude-sonnet-5-5 [Claude Code]

🤖 Generated with Claude Code