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 */
/*
* api_speech.h - the speech model of the public API (api_speech.c) and
* its diarization (api_speech_diar.c): a Whisper model, its voice activity
* detector and speaker model, the last transcription and its turns.
*/
#ifndef JANAS_LLM_API_SPEECH_H
#define JANAS_LLM_API_SPEECH_H
#include "janas/llm.h"
#include "speaker.h"
#include "vad.h"
#include "whisper.h"
/* a speaker's turn: consecutive words of one speaker */
struct janas_turn {
double t0, t1;
char *text;
char *speaker; /* a known speaker's name, or "A", "B"... */
};
struct janas_llm_speech {
struct janas_whisper *w;
struct janas_whisper_result r;
struct janas_vad *vad; /* or NULL */
struct janas_speaker *speaker; /* or NULL */
struct janas_turn *turns;
uint32_t n_turns;
};
/* the model beside path (its directory) named file, opened by open, or
NULL when there is none */
char *janas_speech_beside(const char *path, const char *file);
/* the speakers of the transcription just made (s->r, its words' times)
of pcm (n samples at 16 kHz), from the stretches of speech regions (NULL:
found here, or the whole recording without a detector): its turns in
s->turns. 0, or -1 with the reason in err (also a known speaker's
recording that cannot be read) */
int janas_speech_diarize(janas_llm_speech *s, const float *pcm, size_t n,
const size_t *regions, size_t n_regions,
const struct janas_llm_transcribe_params *p, char *err,
size_t err_len);
void janas_speech_turns_free(janas_llm_speech *s);
#endif