transcribe-cpp-sys 0.0.1

Native FFI bindings for transcribe.cpp, a C/C++ speech-to-text library built on ggml
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59

// arch/funasr_nano/adaptor.h - audio adaptor graph builder.
//
// 2-layer transformer adaptor (FunASR's "Transformer" adaptor class):
//
//   x_in  [encoder_dim=512, T_lfr]
//     -> linear1 [encoder_dim, pre_ffn_dim=2048] + bias  = adaptor.linear1.out
//     -> ReLU
//     -> linear2 [pre_ffn_dim, llm_dim=1024]    + bias   = adaptor.linear2.out
//     -> blocks[0..1]: pre-LN MHA (LayerNorm eps=1e-12) + bottleneck FFN
//                      (1024 → 256 → 1024, ReLU, biases)  = adaptor.blocks.0.out
//                                                            adaptor.out
//
// downsample_rate=1 means linear1 input is encoder_dim (no fold). Only
// retained as a configuration knob for sibling variants that may set k>1.

#pragma once

#include "ggml.h"

struct ggml_context;
struct ggml_tensor;
struct ggml_cgraph;

namespace transcribe::funasr_nano {

struct FunAsrNanoHParams;
struct FunAsrNanoWeights;

struct AdaptorDumps {
    ggml_tensor * linear1_out  = nullptr;  // post-linear, pre-ReLU
    ggml_tensor * linear2_out  = nullptr;
    ggml_tensor * block0_out   = nullptr;
    ggml_tensor * adaptor_out  = nullptr;
};

struct AdaptorBuild {
    ggml_tensor * enc_in  = nullptr;   // [encoder_dim, T_lfr] f32
    ggml_tensor * out     = nullptr;   // [llm_dim, T_lfr]
    AdaptorDumps  dumps {};
    ggml_cgraph * graph   = nullptr;
};

AdaptorBuild build_adaptor_graph(ggml_context *             compute_ctx,
                                 const FunAsrNanoWeights &  weights,
                                 const FunAsrNanoHParams &  hp,
                                 int                        T_in);

// Compute the LLM-injected slice length for use_low_frame_rate=true.
// Mirrors FunASR's data_load_speech formula:
//
//   o1 = 1 + (T - 3 + 2) / 2     // = (T - 1) / 2 + 1
//   o2 = 1 + (o1 - 3 + 2) / 2
//   fake_token_len = (o2 - 1) / 2 + 1
//
// With T=183 → 23. With use_low_frame_rate=false the spliced length
// equals the input frame count.
int compute_fake_token_len(int T_lfr, bool use_low_frame_rate);

} // namespace transcribe::funasr_nano