FireRedVAD (GGUF)

GGUF conversions of FireRedTeam/FireRedVAD (Apache-2.0, © Xiaohongshu FireRedTeam), for use with ggml.

File	Upstream	Context	sha256
`firered-vad-v1.gguf`	`VAD/`	20 frames back, 20 ahead (200 ms look-ahead)	`91b883a92ae7ff498be16f7aa7206e60c7494837d7b395119283faa4b6895eaf`
`firered-stream-vad-v1.gguf`	`Stream-VAD/`	20 frames back, causal	`78db2285c85b662dc350911949fbc90495a52ec14dc0c19998d18fadd8b8f5a4`

Weights are unchanged float32 (bit-exact to the upstream model.pth.tar, HF revision 7990aaccc6b7aec1e527743bd30201f2c4a03b8c), with the original tensor names.
Hyperparameters are stored as firered-vad.{idim,R,M,H,P,N1,S1,N2,S2,odim}.
The Kaldi CMVN from cmvn.ark is stored as cmvn.mean and cmvn.istd (80 each).
Input features: 80-bin Kaldi fbank, 16 kHz, 25 ms window, 10 ms shift, then CMVN. Output: one speech probability per 10 ms frame.

Converted with a standard-library Python script that reads the checkpoint through a restricted unpickler (no code execution). Please cite and credit the upstream authors: FireRedVAD paper.