37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models compared in Q4_K_M. LLMs bigger than your RAM, pure C, CPU or GPU.
C avx2ccpu-inferenceggufinference-enginelinuxllmlocal-llmmixture-of-expertsmoenvmequantizationqwen3simdspeculative-decodingvulkan
/* SPDX-License-Identifier: GPL-3.0-or-later
Copyright (C) 2026 Maurizio Cammalleri */
/*
* vad.h - voice activity detection with Silero VAD (vad.c): the
* probability of speech in every 32 ms of 16 kHz sound, and the stretches
* of speech found from it as Silero's get_speech_timestamps finds them.
*/
#ifndef JANAS_LLM_VAD_H
#define JANAS_LLM_VAD_H
#include <stddef.h>
#include <stdint.h>
struct janas_vad;
/* a model made by tools/silero2jns, or NULL with the reason in err */
struct janas_vad *janas_vad_open(const char *path, char *err, size_t err_len);
void janas_vad_close(struct janas_vad *v);
/* the samples a probability covers (512) and their rate (16000) */
uint32_t janas_vad_window(const struct janas_vad *v);
uint32_t janas_vad_rate(const struct janas_vad *v);
/* the probability of speech in each window of n samples (the last one
padded with silence), into probs (n / window rounded up, malloc'd):
their count, or -1 when out of memory */
long janas_vad_probs(struct janas_vad *v, const float *pcm, size_t n,
float **probs);
struct janas_vad_params {
float threshold; /* 0.5 */
uint32_t min_speech_ms; /* 250 */
uint32_t min_silence_ms; /* 100 */
float max_speech_s; /* 0: no limit */
uint32_t speech_pad_ms; /* 30 */
};
/* the stretches of speech of n samples from their probabilities: pairs
(start, end) in samples into *seg (malloc'd), their count, or -1 */
long janas_vad_segments(const struct janas_vad *v, const float *probs,
long n_probs, size_t n,
const struct janas_vad_params *p, size_t **seg);
#endif